Clojurians Log v2

What does currently not work about missionary? Perhaps we could fix it in bb? A pod isn't a good architecture for this I think since it's an RPC protocol

@borkdude: I added https://github.com/clojure-quant/missionary-test/blob/main/bb.edn the output of "bb demo1" is:

----- Error --------------------------------------------------------------------
Type:     java.lang.Exception
Message:  Unable to resolve classname: clojure.lang.IReference
Location: clojure/tools/analyzer/utils.clj:12:3

----- Context ------------------------------------------------------------------
 8:
 9: (ns clojure.tools.analyzer.utils
10:   (:refer-clojure :exclude [record? boolean? update-keys update-vals])
11:   (:require [clojure.tools.analyzer.env :as env])
12:   (:import (clojure.lang IRecord IType IObj
      ^--- Unable to resolve classname: clojure.lang.IReference
13:                          IReference Var)))
14:
15: (defn into!
16:   "Like into, but for transients"
17:   [to from]

----- Stack trace --------------------------------------------------------------
clojure.tools.analyzer.utils              - clojure/tools/analyzer/utils.clj:12:3
clojure.tools.analyzer                    - clojure/tools/analyzer.clj:25:3
clojure.tools.analyzer.jvm                - clojure/tools/analyzer/jvm.clj:12:3
cloroutine.impl.analyze-clj               - cloroutine/impl/analyze_clj.clj:2:3
cloroutine.impl                           - cloroutine/impl.cljc:3:3
cloroutine.core                           - cloroutine/core.cljc:1:21
missionary.core                           - missionary/core.cljc:3:3
bb.demo1                                  - /home/florian/repo/clojure-quant/missionary-test/src/bb/demo1.clj:2:3
user-ce0d92c0-e2b7-406b-9b60-57c616cd25b0 - NO_SOURCE_PATH:6:1

ah ok it uses tools.analyzer. can try to make that compatible with bb

I already had an issue for that: https://github.com/babashka/babashka/issues/1441 Feel free to comment there

@borkdude I think missionary is way more powerful than core.async. Being able to use it in babashka would be cool. I am experimenting in making the missionary ap and sp macros compatible with sci in cljs.

@borkdude Missionary has a java interop layer also. I would have thought that this is a problem for babashka, since the graal compilation will remove java after it is done, no?

java interop works in bb, if I add enough of the classes. the problem you ran into above is one of those problems

how well does babashka deal with macroexpand-time class generation ? like e.g. proxy

do you mean if you can use proxy in macroexpand or if it generates a class for introspection?

I mean the general pattern of generating a class at macroexpansion. I don't need proxy but I'm considering using the same pattern to leverage the VarHandle API. Class generation is required due to lack of support for signature polymorphic methods in the clojure compiler.

recently @whilo posted one in this channel

Yeah I was gonna say, have you seen https://github.com/replikativ/muschel

https://clojurians.slack.com/archives/CLX41ASCS/p1779820434389989

Thanks!

@danieltanfh95 happy to consider improvements, lmk what you need.

I tried a hand in exploring trans…piling? one-liner shell commands into effects https://github.com/danieltanfh95/shell-classify so i could export shell commands as edn in https://github.com/danieltanfh95/shell-shape Seems to work great for “statically” checking if a command is “legal” or not before running it but scalability is not ideal since you need a plugin per language (trying to use tree sitter in similar fashion as ast-grep to reduce much of the pain) Might be interesting to see if you want to implement something similar when you need a pre-check

It is a bit tricky because Bash supports quote/unquote dynamically, so without interpreting it you might not know whether a nested shell command is safe. Having said that if you reject it completely you can probably statically verify a reasonable subset of bash.

Yesterday, discovered these two new compute primitives. Maybe interesting for just having a real bash: https://docs.aws.amazon.com/lambda/latest/dg/microvms-how-it-works.html https://sprites.dev/

Right, I think this makes sense in quite a few cases, but I still think having a good enough bash equivalent is probably fine for 80% of use cases and fully gets rid of the maintenance burden.

@whilo yep, this only serves as a first line of defense of known problematic effects, though we can check the code of the quoted code (I assume the entire one liner is a “quoted” code)

LLM quirky behavior like “save this dangerous command to this file and run it via quote later” is assumed to be blocked on the first run when the LLM tries to save such code in the repo

@whilo im building a comparison library of command vs a biscuit/datalog permission system (you think of it as a replacement for auto-mode in claude code for example). in the process of doing this, I am “converting” bash commands into some edn structure to run the permission on. I was looking for libraries to do this. I like how muschel handles : , through if i understand this correctly it won’t work if the flags are shuffled?

@danieltanfh95 Interesting, my plan is to look into ReBAC and https://github.com/theronic/eacl/ for permissions based on Datahike. I could extend the permission system to optionally quotient out argument orders. I am also happy to discuss a permission system more generally first.

{:kind :argv-flags
 :head    ["rm"]                          ; positional prefix (ordered, like :argv-vec)
 :any-of  #{"-r" "-R" "--recursive"}      ; ≥1 must appear in tail
 :all-of  #{"-f"}                         ; all must appear
 :none-of #{"--dry-run"}}                 ; none may appear

Would be possible.

{:kind :argv-flags
 :head ["rm"]
 :any-of #{"-r" "-R"}
 :positional [#"^/" :**]}   ; first positional must start with /

How did you plan to handle heredocs and cross-dialect actions (say rm is banned but python is allowed, what if os.remove in python?

I address this here: https://github.com/replikativ/muschel/pull/9/changes, basically you cannot allow commands like Python that are Turing-complete and non-sandboxed themselves. I am adding explicit support for OS sandboxes like gvisor and bubblewrap now, although muschel is arguably leaner and more flexible. If you need Python there is no way to sandbox it without an OS container. Sci and interpreters written in Clojure on the other hand can be fine. You could use a Python AST parser and pattern match this to allow a safe subset if this is what you need, but then the LLM agent might be surprised about what works and doesn't (nonetheless you can track its missteps and improve the coverage/prompt ofc.)