Tags

Squid 1.8.6

1.8.5

- Harden in-flight script ledger: bounded lock stripes + retried ensure-row (#386)
- Add Calamari script wall-clock timeout; localize K8s namespace resolution out of the generic render path (#387)
- Add crash-resume re-attach integration coverage for the resume-by-ticket wiring (#388)

Squid 1.8.4

Resume in-flight scripts by ticket — a server crash mid-deployment re-attaches
to the still-running agent script instead of double-dispatching. Deployment
capability enforcement (off/warn/strict, default warn) on a target plan unified
as the single source of truth for both preview and execution, so the machines a
preview reports as matched are exactly the machines the deployment runs on.

Squid 1.8.3

Multi-shell Calamari (PowerShell + Python runners), .properties/.ini plus
YAML/XML structured-config formats, .7z package extraction, multi-shell
convention hooks (PreDeploy/PostDeploy/DeployFailed), and environment
rollback (deployment-journal target resolver + rollback command).

1.8.2 — Calamari G1.x rewriter pipeline + multi-format + multi-shell + structured outcomes

Squid 1.8.1 — 1.8.0 hardening review followups

Post-merge self-audit of the 1.8.0 Upgrade Hardening initiative
surfaced one cosmetic bug and three test-coverage gaps that wouldn't
have caught a future regression. This release addresses all four in
PR #362.

  Bug fix (H4) — Clock-skew Math.Max(0, ...) guard for `elapsed` in
  upgrade-lock contention message. Pre-fix, two server pods disagreeing
  on UTC (NTP slew, fresh-pod startup) showed the operator "~-3599s
  ago" cosmetically; dispatch itself worked correctly.

  Test gap (H7) — CapabilityKeys.Role.* constant pinning added to
  CapabilityKeysConstantsTests (matching Shell / Bin / Privilege
  pattern). Drift in role:iis / role:docker / role:nginx / role:systemd
  literals now surfaces at build time.

  Test gap (H7) — TentacleHealthCheckStrategy cache-population test
  with installedRoles in metadata. Pins the read-side wire contract
  between agent capability emission and server cache write.

  Test gap (H7) — MachineRuntimeCapabilitiesPersistence JSON shape
  tests pin installedRoles serialisation, deserialisation, AND the
  1.8.0 → 1.8.1 backward-compat path (blobs without installedRoles
  field deserialise to null).

5603/5603 unit suite green (+7 net new vs 1.8.0).

No operator-visible behaviour change beyond the cosmetic contention
message fix.

Squid 1.8.0 — Upgrade Hardening initiative

8 PRs (H1-H8) closing every link in the 1.7.x operator-facing failure chain
discovered during the Windows IIS deploy field test.

  H1 (#354) — UpgradeEligibilityReason codes + cold-cache NoOsDetected
              + OS-aware registry-unreachable messages. Windows operators
              no longer see misleading "Docker Hub unreachable" hints.

  H2 (#355) — Capability cache persisted to machine.runtime_capabilities_json.
              Server pod restart no longer wipes the cache.

  H3 (#356) — Active health-check returns structured ManualHealthCheckResult
              with fresh AgentVersion/Os + structured ErrorCode.

  H4 (#357) — Upgrade-lock contention message enriched with dispatchedAt +
              targetVersion + expected-remaining time.

  H5 (#358) — LinuxTentacleUpgradeStrategy strict-Linux-only; UpgradeAsync
              cold-cache guard. Direct API can no longer bypass H1.

  H6 (#359) — MachineUpgradeStatus.RolledBack distinct from Failed. Operators
              see "safely restored to baseline" instead of "broken".

  H7 (#360) — role:* capability slots. IIS deploy to non-IIS machine fails
              at plan-time with actionable hint, not at runtime with
              "Import-Module WebAdministration: module not found".

  H8 (#361) — Composition regression-guard tests pinning the cumulative
              H1-H7 behaviour. 11 new tests; 5596/5596 total unit suite
              green; +79 net new tests across the initiative.

Operator-facing impact: every failure mode encountered in the 1.7.x chain
is now structurally impossible. Cross-OS upgrade is uniform; failed
upgrades self-recover via rollback (.bak / apt-snapshot); plan-time
validation catches missing target capabilities before any wasted dispatch.

Squid 1.7.9

Bug fix: Skip Calamari for PowerShell scripts (was crashing with
FileNotFoundException).

Tentacle's StartProcess routed PowerShell + variables.json through the
Calamari child process, but Calamari has no PowerShell pipeline:
WriteBootstrappedBashScriptStep does File.ReadAllText("script.sh") and
the CLI is hardcoded to --script=script.sh. Result on every Windows IIS
deploy: FileNotFoundException → Calamari child died → state file stuck
at Progress=Running → server's GetStatus orphan-detection returned
Complete + UnknownResult (-1). Operator saw "Unknown result (ticket or
process not found) (exit code -1)".

Fix: IsCalamariCompatible(syntax) returns true ONLY for Bash. PowerShell
scripts now route through StartViaLauncher, which picks the right
launcher (pwsh-Core if installed, Windows PowerShell 5.1 fallback per
1.7.8's PR #352). Safe because all server-side PowerShell builders
already inline variables into the rendered script body — Calamari's
bootstrap preamble was adding zero value on top of that.

Squid 1.7.8

Bug fix: Fall back to Windows PowerShell when pwsh-Core not installed.

ProcessLauncherFactory unconditionally defaulted to pwsh (PowerShell 7
Core) on Windows, but pwsh is OPTIONAL — Microsoft doesn't bundle it
with the OS. Every stock Windows tentacle without PS7 installed crashed
with Win32Exception (2) "system cannot find the file specified" the
moment any PowerShell script dispatched (upgrade, deployment, health
check). Fix: probe pwsh.exe availability and auto-fall-back to the
always-present OS-bundled PowerShell.exe 5.1 when absent.

Operator opt-in via SQUID_TENTACLE_USE_WINDOWS_POWERSHELL=true still
wins over the probe.