stephrobert/feint v0.16.0

on GitHub

Added

• A stack preflight reads what a configuration declares, before anything
  starts (#888). feint run reads the Terraform or OpenTofu stack its child
  would read and refuses, before the lab is up, a configuration declaring a type
  a pack refuses; feint doctor --stack <dir> asks the same question outright,
  changing nothing. --preflight takes auto (the default, refuses only a
  declined type), strict (also refuses an unmeasured one) and off;
  --preflight-json and feint doctor --stack-json write the report where CI
  can read it. A refusal is exit 1: 2 is drift, and the codes are a contract.

  What it replaces, measured on 2026-11-12 with provider outscale/outscale
  1.8.0: a stack with three declined types applied 48 resources, failed with two
  404s, left 50 entries in the state, and never attempted the third declined
  type at all, because its dependencies had failed — so fixing the two visible
  errors bought a second full apply to find the third. The preflight names all
  three in 12 ms, with nothing created.

  compatible means a recorded run applied and destroyed those types against
  this emulator. It is not a promise about every attribute, nor about anything
  past the control plane, and a type whose product is outside the drift scan
  reads as unknown rather than refused. docs/limits.md says what the reading
  cannot cover.

• feint checkpoint and feint reset (#853): a checkpoint records the
  control plane and the object bytes under one name, and a reset returns the
  lab to it — destroying what the checkpoint does not hold, starting again what
  the host no longer holds, and putting each pack's firewall rule sets back on
  the runtime. The verdict is not the operation's own: it compares what a
  client can observe, control plane, bytes read back through the doors, and
  what the host holds, and it says did not converge rather than reporting the
  intention.

• A Linux install route in the Quick Start (#843), under the Homebrew one
  and generated from the same version the install page is pinned to. Six lines,
  and they verify: the bytes are checked against the published checksums.txt
  before anything is installed. What that does not prove is said rather than
  implied — it proves the bytes match the list, not who wrote the list, and the
  signature over the list stays in docs/install.md.

• feint run owns the lab (#848): with no --endpoint it starts an
  emulator, waits for it, resolves the provider environment the way feint env
  does, runs the command and stops everything it started. --provider narrows
  the exports to one pack, --keep-on-failure leaves a failed run's lab up and
  prints the exact command to reach and remove it, --json <path> writes the
  receipt where it is asked for, and --runtime asks the host for a machine
  runtime before anything starts, with the refusal feint up makes. The exit
  code is the child's, and stdout belongs to the child alone.

• A run removes the variables a pack measured as a way back to the real
  cloud, rather than only overriding them: OSC_PROFILE makes the Outscale
  provider 1.1.x read ~/.osc/config.json and ignore the endpoint export
  entirely (#342), so overriding does not help and the variable has to go. The
  names are discovered from the packs' own EnvHazards, not kept in a second
  list here.

Changed

• The Outscale example stack pins its provider exactly (#873, #898), at
  1.8.1, where it floated at ~> 1.7. docs/clients.md's row for it goes
  from "resolved fresh on each run, so the version that answered is not knowable
  here" to "exact: the version that answered". The floor was right and the
  floating part was the accident #778 measured on another stack: two engines
  resolving one constraint to two different versions, both reporting success.
  The module's own ~> 1.7 stays, deliberately — it is a floor for reuse, and
  Terraform intersects it with the root's pin.

• The API descriptions are re-extracted, and Outscale's document moved from
  1.42.0 to 1.43.2 (#900), 655 schemas to 656: RequestedSnapshotMode
  added, Snapshot, FiltersSnapshot, AccessKey, ApiAccessPolicy and
  UpdateApiAccessPolicyRequest changed. Scaleway's baremetal options moved
  with them, gaining private_network, public_bandwidth, remote_access,
  license and certification. No newly declared query parameter lands on a
  route this emulator mounts, which is the case that would have cost code.

• The coverage headline names the population it counted. It read "Of the
  548 operations upstream declares" while the Scaleway scan covers 11 of the 56
  products in that SDK, so a third of the provider was printed as the provider.
  A stack using a product outside the scan reaches the real endpoint, and the
  README now says which scan it is talking about.

• README.md stops saying object storage is out of scope. It shipped in
  0.15.0; the sentence sat outside the generated blocks, where feint docs --check could not reach it.

• A tag is refused while an open milestone claims a version already tagged
  (#877). v0.14.0 and v0.15.0 were both published while milestones named
  after them stayed open with 19 and 10 issues, so the tracker said two
  released versions were unfinished. tools/release/milestone-gate.sh asks
  before each tag, and says when it could not measure rather than refusing for
  a reason that is not about the repository.

• docs/limits.md caught up with behaviour that changed months ago (#857).
  Eleven sections cited issues that had closed since anyone last read them;
  three were false, including a heading saying the Exoscale Terraform provider
  is refused when it has been driving the pack since September.

• The boundary of this product is written down (#196, #733):
  docs/limits.md states that feint never manages a resource whose loss would
  matter. Two current texts had said opposite things for two months, neither
  citing the other.

• feint run leaves stdout to the command it runs. Its own prose goes to
  stderr and the JSON receipt to a path. Measured rather than reasoned: the
  first version wrote the receipt to stdout, and replaying a surveyed stack
  made terraform apply print its plan into the same stream, so the receipt
  was unparseable exactly when the run was real.

• A run refuses a measured escape it has no door for, before the child
  starts. Not every hazard: the ones the packs report are what this command
  intercepts, so the question is whether THIS run mounts a door for the pack
  that reported it. The receipt also reports the runtime and the capabilities
  the emulator declares, rather than the mode that was asked for.

Fixed

• Six query parameters a served route declared and dropped in silence
  (#884, #875). The Scaleway API description had not been re-extracted since
  2026-10-19 while the documents moved under it; a fresh extraction answered
  +328 lines, holding twenty-one newly declared query parameters, six of them on
  routes this emulator serves. server, detached, regional, image_id,
  version_id and security_group_ids are now served, or refused by name.
  version_id is refused before anything else is resolved, so a request
  carrying it with an unknown label is told which argument is the problem rather
  than answered "no such image". Found by an outside contributor, as a side
  effect of a triage.

• The nightly drift scan reports a debt, not only an arrival (#892, #886).
  The gate asked whether something appeared that is not in the committed
  baseline, and the baseline records the unknown operations — so the first
  refresh after an arrival wrote it down and every later night found nothing
  new while the operation was still undecided. Measured: two baremetal
  commitment operations sat undecided for fourteen green nights. Three more
  halves of the same mechanism moved with it: the pull request now carries the
  API descriptions it regenerated, which had not travelled since 2026-10-19; the
  triage view covers the eleven products the gate scans rather than three, so an
  operation outside those three can now be named; and a source that could not be
  measured makes the night red instead of reading as an absence of drift.

• The drift report finds its own issue by title, not by label alone (#894).
  The lookup was --label drift plus the newest match, and three open issues
  wear that label. With the backlog empty, the first quiet night would have
  closed the issue carrying the work still to do, with the words "Every
  operation upstream is triaged".

• The falsification replay fits its budget, and a ceiling it reached is a
  measured fact (#893, #890). The scheduled run of 2026-11-12 was cancelled
  after 210.3 minutes against the timeout-minutes: 210 it declares for itself,
  and the issue it opened called that the run's own infrastructure. The cost was
  in the repository: instrumented on one spec, the two full test suites it runs
  per spec were 45 of 53 seconds, which across a replay is 242 baselines and 242
  restorations for some fifteen distinct questions. The baseline is now answered
  once per packages and environment, and the restoration is read off the bytes
  of the files that were mutated. Eight ./internal/cli/ specs, 54 mutations,
  every one still biting: 441 s before, 181 s after. The night report tells a
  budget reached from a runner lost, and names the duration and the ceiling.

• A test asked the kernel for a promise it never made (#896). The free-port
  probe closes its listener before answering, so the same port may be handed
  straight back; a test required two consecutive probes to differ. Measured over
  two thousand pairs on one station: one came back identical. It asserts
  loopback and bindable now, which is what the function promises — its own
  comment opens with "Racy by nature".

• A reset puts a firewall rule back on the interface, not only in the store
  (#871). Survey answers names, and a rule set whose content changed keeps its
  name, so nothing in the observation could see the difference: a reset reported
  convergence while the runtime went on enforcing the rules of a state the
  checkpoint does not describe. Reproduced with a TCP connection under
  --vm incus before being fixed. The same pass fixed a reset starting the
  machine of a record the checkpoint describes as stopped, and then writing
  running over the state it had been asked to return to.

• The reset verdict compares the bytes of an object, not just its name
  (#871). It read the digest from a field ListObjectsV2 does not emit, so the
  field was always empty and two different contents under one key compared
  equal. Each object is now fetched and hashed by the observer.

• feint run fails when it did not keep a promise (#871). Only
  did not converge used to fail the run, so a failed reset, one that could not
  be read back, and a lab this command owned and could not stop all left the
  exit code at the child's.

• A checkpoint is whole or it is the previous one (#871). The bytes were
  saved before the control-plane snapshot, so a failure in the second step left
  an earlier checkpoint of that name holding its old store beside the new object
  references.

• Named checkpoints stop overwriting each other (#871): the emulator holds
  any number and the CLI wrote one file per instance, so checkpoint A then
  checkpoint B made reset A impossible.

• The gofmt pre-commit hook fails when gofmt cannot run (#829). It read
  the list of offending files and never the exit code, so any failure of gofmt
  produced an empty list and the hook reported Passed. Its window is a fresh
  clone before mise trust, which is a new contributor's first commit.

• A balancer verdict is drawn from twenty-four connections, not six (#559).
  Six is a coin flipped six times: a balancer that works sent every connection
  to one side about one run in thirty-two, so the dataplane suite reddened on a
  subject that was behaving. The same arithmetic ran the other way after an
  unlink, as a false green over a real defect.

Proved against

The clients

The Terraform providers

The client above is the engine; what answers this emulator is the provider.
Each row is one required_providers entry, read where it is written.

Read the third column narrowly, because it is the one a consumer pins against.
An exact constraint names the version that answered. A constraint is
resolved by terraform init -upgrade on the runner, every run, so the version
that answered is not knowable from this repository — it is a floor, not a
proof. Not pinned is not pinned: nothing here says which version ran, and
no number is invented to fill the cell.

Every stack under examples/stacks/ is applied on every pull request.

Generated from the conformance workflow and the fixtures at this tag.
A newer client is not known to work until the pin moves and the suite
passes on it: docs/clients.md says what each cell is worth.