feat: lua script commands triggers - #447

This PR doesn't contain any Lua.

The branch is feat/lua-script-commands-triggers and the description walks through the Lua triggers, the properties table, and 36 passing tests, but the diff is:

AGENTS.md
plugin-script-perl/...  (7 files)
plugin-script-r/...     (7 files)

Zero files under plugin-script-lua. The two commits on the branch are feat/r-script-commands-triggers and feat/perl-script-commands-triggers. plugin-script-lua on main still has only Script.java, Commands.java and package-info.java, so the work genuinely isn't here.

Separately, this diff is byte-identical to #446 (both 1950 lines), and #438 is the R subset of both. So #438, #446 and #447 are three PRs carrying overlapping copies of the same commits rather than one module each. Merging any one of them will conflict the other two.

Could you rebase each onto current main with only its own module: #438 R, #446 Perl, and this one Lua? I'll review them properly once they're separated. Nothing here reflects on the code itself, I just can't review Lua triggers that aren't in the diff.

Hey @Abhishek84313 👋, are there any updates on this?

Hey @Abhishek84313 👋, are there any updates on this?

Haven't checked it, i will let you know

Thanks for catching this @Malaydewangan09, you were right on all counts.

The cause was a one-character branch name mix-up on my side: the actual Lua
work was sitting on my local feat/lua-script-commands-trigger (singular),
while the branch I pushed was feat/lua-script-commands-triggers (plural),
which only ever had the R and Perl commits. Hence the identical diffs across
#438/#446/#447.

I've rebased this branch onto current main. It is now a single commit
touching only plugin-script-lua:

• ScriptTrigger.java, CommandsTrigger.java
• ScriptTriggerTest, CommandsTriggerTest, ScriptTriggerConditionTest,
  CommandsTriggerConditionTest, EdgeStateTest

No AGENTS.md, no R, no Perl, and no build.gradle change. Merges cleanly.

Since R (#438) and Perl (#446) are already merged, those two are resolved and
this is the only one left from the set.

While rebasing I noticed my Lua code predated the review feedback on #438, so
I ported both triggers from the merged R implementation rather than ship the
older shape. Three things came across:

• Edge state moved to the namespace KV store. It was previously tracked in
  an AtomicBoolean field, which does not survive a poll: the trigger is
  rebuilt from the flow definition and serialized to a worker each time, so
  edge: true would have fired on every matching poll instead of only on the
  transition. EdgeStateTest now covers this across fresh instances.
• Bounded LRU cache for compiled condition patterns, instead of
  recompiling the regex on every poll.
• Blank exitCondition now throws rather than defaulting to "", which
  previously matched everything.

compileJava and compileTestJava pass, and 41/41 non-Docker tests pass
locally. The Docker-backed cases in CommandsTriggerTest need CI, as I have
no Docker on this machine.

On the issue's proposed commandPattern / argumentsPattern: this PR still
implements only exitCondition, matching every existing reference
implementation. These triggers run the script themselves rather than observing
externally-launched executions, so there is no command stream to match
against. If those properties are wanted, they should land across all submodules
in a separate change.

Thanks, this is the right diff now. Only plugin-script-lua, and both triggers are a clean port of the merged R ones (24 lines differ, all names, image and examples).

Ran it on current main with Docker: 49/49 pass, including the container-backed CommandsTriggerTest cases. Also ran the edge cycle through evaluate() against real containers, trigger reserialized every poll:

cmds : exit 1  exit 1  exit 0  exit 0  exit 1
got  : FIRED   quiet   quiet   quiet   FIRED

Two things missing that the R and Perl PRs did include:

• AGENTS.md still lists only Commands and Script under plugin-script-lua. Please add CommandsTrigger and ScriptTrigger.
• src/main/resources/doc/io.kestra.plugin.scripts.lua.md has no ## Triggers section. The R doc has one, same shape works here.

One optional: shouldEmit writes the KV entry on every poll even when the result hasn't changed. The PowerShell trigger (#436) only writes when matched != previouslyMatched, which cuts a steady condition from ~1440 writes a day to one per transition. Fine to leave if you'd rather stay identical to R.

Happy to approve once the two docs bits are in.

Two more from a closer pass, both small:

• ScriptTrigger.evaluate() is never run against a real container in this module. ScriptTriggerTest and ScriptTriggerConditionTest only call matchesCondition, and EdgeStateTest calls shouldEmit directly. Only CommandsTriggerTest goes through evaluate(), so the Script task wiring (lua <file> in nickblah/lua, TriggerRunContext) has no coverage. A ScriptTriggerEvaluateTest like the PowerShell one would cover it: os.exit(1) with exitCondition: exit 1 asserting exitCode 1, plus one print('::{"outputs":{"k":"v"}}::') case for the vars path.
• The exitCondition docs still say the regex is matched against failure logs (ScriptTrigger.java:108 "emitted vars and failure logs", CommandsTrigger.java:112 "vars + logs"). Logs are never read. buildHaystack is only vars.toString(), and a failed run has no vars, so only exit N can match a failure. The PowerShell wording is accurate: "matched against emitted vars; a failed run has no vars, so only 'exit N' can match a failure."

Thanks @Malaydewangan09, all addressed in 8521610:

AGENTS.md: added CommandsTrigger and ScriptTrigger under plugin-script-lua.
Docs: added a ## Triggers section to io.kestra.plugin.scripts.lua.md, same shape as R.
exitCondition wording: fixed in both triggers. It no longer mentions logs, and now says the regex matches emitted vars only, so only exit N can match a failure (PowerShell wording).
ScriptTriggerEvaluateTest: runs evaluate() against nickblah/lua for os.exit(1) + exit 1 (expects exitCode 1), print('::{"outputs":{"k":"v"}}::') (expects vars {k=v}), a non-matching run, and edge suppression across a serialized and rebuilt trigger.
Optional KV write: took it. shouldEmit now writes only when matched != previouslyMatched, with the two PowerShell no-rewrite tests ported to EdgeStateTest.
43/43 non-Docker tests pass locally. I still have no Docker, so the container-backed ScriptTriggerEvaluateTest / CommandsTriggerTest need your run or CI. The branch still merges cleanly onto current main.

All four addressed in 8521610, plus the write-skip. Ran it with Docker on current main: 55/55, and the ScriptTrigger edge cycle through evaluate() gives FIRED quiet quiet quiet FIRED across exit 1/1/0/0/1. Good from me.