or: fpath=(/path/to/stryke/completions $fpath) in .zshrc

 ███████╗████████╗██████╗ ██╗   ██╗██╗  ██╗███████╗
 ██╔════╝╚══██╔══╝██╔══██╗╚██╗ ██╔╝██║ ██╔╝██╔════╝
 ███████╗   ██║   ██████╔╝ ╚████╔╝ █████╔╝ █████╗
 ╚════██║   ██║   ██╔══██╗  ╚██╔╝  ██╔═██╗ ██╔══╝
 ███████║   ██║   ██║  ██║   ██║   ██║  ██╗███████╗
 ╚══════╝   ╚═╝   ╚═╝  ╚═╝   ╚═╝   ╚═╝  ╚═╝╚══════╝

[THE FASTEST DYNAMIC LANGUAGE IN THE WORLD FOR PARALLEL OPERATIONS]

"There is more than one way to do it — in parallel."

"100% TDP — beware."

"The hottest language ever created. Literally."

[PATENT PENDING]

The 2nd fastest dynamic language runtime ever benchmarked for singlethreaded — behind only Mike Pall's LuaJIT, and beating it on 3 of 8 benchmarks. The fastest on all mulithreaded benchmarks. A Perl 5 compatible interpreter in Rust with native parallel primitives, NaN-boxed values, three-tier regex, bytecode VM + Cranelift JIT, streaming iterators, and rayon work-stealing across all cores. Faster than perl5, Python, Ruby, Julia, and Raku on every benchmark.

Read the Docs · Full Reference

---

Table of Contents

• [0x00] Overview
• [0x01] Install
• [0x02] Usage
• [0x03] Parallel Primitives
• [0x04] Shared State (varsync)
• [0x05] Native Data Scripting
• [0x06] Async / Trace / Timer
• [0x06b] AOP — Before / After / Around Advice
• [0x07] CLI Flags
• [0x08] Supported Perl Features
• [0x08a] --no-interop Mode
• [0x08b] String Coordinates — Bytes vs Codepoints
• [0x08c] --static Mode
• [0x09] Architecture
• [0x0A] Examples
• [0x0B] Benchmarks
• [0x0C] Development & CI
• [0x0C-test] Test Runner — Worker Pool Architecture
• [0x0D] Standalone Binaries (stryke build)
• [0x0E] Inline Rust FFI (rust { ... })
• [0x0F] Bytecode Cache (rkyv)
• [0x10] Distributed pmap_on / ~d> over SSH (cluster)
• [0x10a] Infrastructure Load Testing
• [0x10b] Agent/Controller Architecture
• [0x10c] Scriptable Distributed Compute — congregation / pray / annex
• [0x11] Language Server (stryke lsp)
• [0x12] Language Reflection
• [0x14] Package Manager
• [0x15] Web Framework (s_web)
• [0x16] AI Primitives
• [0x17] Expect / Interactive Automation
• [0x18] Documentation
• [0xFF] License

What's new — NAT traversal + value lineage + multi-target teleport

Stack added across v1 → v1.5 (under [0x10b] Agent/Controller):

Layer	Builtins	What it does
Network probes	kick / udp_send	TCP liveness probe + fire-and-forget UDP send (Wake-on-LAN, NAT keepalive)
P2P primitives	udp_open / udp_send_to / udp_recv / udp_recv_from / udp_close / stun / stun_classify / punch	Persistent UDP socket pool, RFC 8489 STUN client (IPv4 + IPv6), UDP hole-punching, symmetric-NAT detection
TURN fallback	turn_allocate / turn_permission / turn_send / turn_recv / turn_refresh	RFC 8656 TURN client (HMAC-SHA1 auth) for when hole-punching fails (~20-30% of cases: symmetric NATs, UDP-blocking firewalls)
Orchestration	ice::connect (stryke source)	Three-rung ladder: direct → punch → relay, first-success wins
Value lineage	mark / provenance / unmark	Arc-keyed per-value lineage tracking with weak-ref GC, zero overhead when unused
Multi-target IPC	teleport / arrive	POSIX-SHM broadcast value to N stryke processes + per-receiver UDS notify. One segment, N readers — beats N×bincode-over-pipe for big payloads

Demos: p2p_chat.stk / p2p_chat_v2.stk / turn_relay_chat.stk / ice_orchestrator.stk / turn_health_check.stk / provenance_basics.stk / provenance_audit_log.stk / provenance_chain_walkthrough.stk / teleport_broadcast.stk

---

[0x00] OVERVIEW

stryke parses and executes Perl 5 scripts with rayon-powered work-stealing primitives across every CPU core. Highlights:

• Server farms first — the first language designed for distributed infrastructure load testing
• Bare metal heat — heat(60) pins ALL cores to 100% TDP for 60 seconds
• Agent/Controller architecture — stryke controller + stryke agent for fleet-wide stress testing
• AI is a primitive, not a library — ai "summarize this", $doc with auto-attached tool fns, MCP client+server, agent loop, RAG memory, vector search (§ 0x16)
• Web framework s_web — Rails-shaped scaffold + ERB engine + SQLite ORM + admin panel + auth + PWA + Dockerfile, s_web new app --app everything --theme cyberpunk --auth --admin --migrate (§ 0x15)
• PTY-driven interactive automation — pty_spawn/pty_expect/pty_send/pty_interact, the modern Tcl/Expect successor with cluster fanout (§ 0x17)
• rkyv KV store — world-first. First-class CRUD store with rkyv as the on-disk codec. kv_open/kv_put/kv_get/kv_del/kv_exists/kv_keys/kv_scan/kv_len/kv_commit/kv_batch/kv_close/kv_stats ship as native builtins. Reads are mmap + validate + cast (zero-copy, no parse, no allocate per row) — the same primitive script_cache.rs already uses for bytecode. SQLite-shaped API ergonomics, beats SQLite on reads for any store that fits comfortably in RAM. Atomic rewrite on commit (tmp + rename), versioned format header (STKV magic + format_version), all-or-nothing kv_batch. Phase 2 ships stryke kvd server + remote kv_connect client over the same archive bytes.
• Sketch algebra — world-first. Probabilistic data structures are first-class operands for + | & ^ -: $bloom_a + $bloom_b (Bloom union), $hll_a + $hll_b (HyperLogLog union), $cms_a + $cms_b (Count-Min counter sum), $topk_a + $topk_b (SpaceSaving merge), $td_a + $td_b (t-digest centroid merge), $rb_a | $rb_b / & ^ - (Roaring set algebra). Operators are functional — operands are never mutated. No other language treats these as syntactic primitives; everywhere else they are library function calls.
• All zsh glob qualifiers in a scripting language — world-first. Every qualifier from zsh's zshexpn(1) works wherever stryke takes a glob (glob, glob_par, slurp/c/cat, swallow/swa, ingest/ing, pwatch, <...>, par_find_files): file-type, permission, ownership, size/links/time numerics, sort + descending sort, [N,M] selection, (N) null-glob, (D) dotfiles, (F) non-empty dir, (f<bits>) mode match, (d<N>) device, (e'CMD') eval, (P…)/(Q…) join words, ^ negate, - follow-symlinks toggle, , OR, : colon modifiers. Backed by the zshrs glob engine — single source of truth, zero stryke-side reimplementation.
• Package manager — Cargo-shaped stryke.toml + stryke.lock, s add/s install/s tree resolver, hash-pinned reproducible builds (§ 0x14)
• New Parallel Subroutines and |> Pipeline Syntactic Sugar
• Runtime values — StrykeValue is a NaN-boxed u64: immediates (undef, i32, raw f64 bits) and tagged Arc<HeapObject> pointers for big ints, strings, arrays, hashes, refs, regexes, atomics, channels.
• Three-tier regex — Rust regex → fancy-regex (backrefs) → pcre2 (PCRE-only verbs).
• Bytecode VM + JIT — match-dispatch interpreter with Cranelift block + linear-sub JIT (strykelang/vm.rs, strykelang/jit.rs).
• Rayon parallelism — every parallel builtin uses work-stealing across all cores.
• 10,453 standard library primaries in %b (11,229 keys in %all including aliases and keywords) — largest bareword library of any language; clears Wolfram v14.3's high-band estimate (~7,300) by ~3,149
• 44 MB single static binary — ~/.cargo/bin/s ships every builtin in one file, ~4.3 KB amortized per builtin, ~200× denser than Wolfram Engine per builtin/byte, sub-10 ms cold start

---

[0x01] INSTALL

# Via Homebrew tap (auto-bumped by each release; formula is `stryke`)
brew tap MenkeTechnologies/menketech
brew install stryke

# Or via crates.io
cargo install strykelang

# Or from source
git clone https://github.com/MenkeTechnologies/strykelang && cd strykelang && cargo build --release

Zsh tab completion

cp completions/_stryke /usr/local/share/zsh/site-functions/_stryke
# or: fpath=(/path/to/stryke/completions $fpath) in .zshrc
autoload -Uz compinit && compinit

stryke <TAB> then completes flags, options, and script files.

---

[0x01b] CONCISENESS — STRYKE VS THE WORLD

stryke is the most concise yet readable ASCII-only general-purpose scripting language — shorter than Perl, Ruby, Python, and AWK for real-world tasks.

vs mainstream languages

Task	stryke	chars	perl	chars	ruby	chars	python	chars
hello world	p"hello"	8	print"hello"	12	puts"hello"	10	print("hello")	14
sum 1-100	p sum 1:100	11	use List::Util'sum';say sum 1..100	38	p (1..100).sum	15	print(sum(range(1,101)))	24
double+filter+sum	~>1:10map{_*2}fi{_>5}sum p	28	say for grep{$_>5}map{$_*2}1..10	36	p (1..10).map{...}.select{...}.sum	42	print(sum(x for x in[...]))	56
max of list	p max 3,1,4,1,5	15	use List::Util'max';say max(...)	38	p [3,1,4,1,5].max	17	print(max([3,1,4,1,5]))	23
reverse string	p rev"hello"	12	say reverse"hello"	18	puts"hello".reverse	18	print("hello"[::-1])	20
count array	p cnt 1:10	10	say scalar 1..10	17	p (1..10).count	16	print(len(range(1,11)))	23
join with comma	p join",",1:5	14	say join",",1..5	17	puts (1..5).to_a.join(",")	24	print(",".join(map(...)))	36
first element	p first 1:10	13	say((1..10)[0])	16	p (1..10).first	16	print(list(range(...))[0])	27
any even	p any{even}1:5	14	use List::Util'any';say any{$_%2==0}1..5	42	`p (1..5).any?{	x	x%2==0}`	25
unique values	p uniq 1,2,2,3	15	use List::Util'uniq';say uniq(...)	38	p [1,2,2,3].uniq	17	print(list(set([...])))	27

stryke wins every task against Perl, Ruby, and Python.

vs K (array language)

K is more terse for pure array math: +/1+!100 (8 chars) vs stryke p sum 1:100 (11 chars). But K is a financial DSL, not a general-purpose language — it lacks:

Feature	stryke	K
HTTP client	fetch"url"	❌
JSON parsing	json_decode $s	needs lib
Regex	$s=~/\d+/	limited
SHA256/crypto	sha256"data"	❌
Parallel map	pmap{$_*2}@a	❌
Compression	gzip $data	❌
Base64	b64e"hi"	❌
UUID	uuid	❌
SQLite	db_query $db,$sql	❌
TOML/YAML	toml_decode $s	❌

K is a calculator. stryke is a programming language.

vs golf languages

GolfScript, Pyth, 05AB1E, Jelly — these are shorter but are write-only puzzles designed for competitions, not real software. stryke remains readable and maintainable.

---

[0x01c] WHY STRYKE — ONE-LINER COMPARISON

stryke is a one-liner-first language. No -e flag needed, everything built in, shortest syntax wins.

Character count — real tasks

Task	stryke	perl	ruby	python	awk / other
Print hello world	s 'p "hello world"' 19c	perl -e 'print "hello world\n"' 32c	ruby -e 'puts "hello world"' 29c	python3 -c 'print("hello world")' 34c	echo | awk '{print "hello world"}' 36c
Sum 1..100	s 'p sum 1..100' 16c	perl -MList::Util=sum -e 'print sum 1..100' 45c	ruby -e 'puts (1..100).sum' 28c	python3 -c 'print(sum(range(1,101)))' 38c	—
Word frequencies	s -an 'freq(@F) |> dd' 22c	perl -ane '$h{$_}++ for @F}{print "$_ $h{$_}\n" for keys %h' 61c	—	—	awk '{for(i=1;i<=NF;i++) a[$i]++} END{...}' 65c+
SHA256 of file	s 'p s256"f"' 13c	perl -MDigest::SHA=sha256_hex -e '...' 70c+	—	python3 -c 'import hashlib;...' 80c+	shasum -a 256 f 15c
Fetch JSON API	s 'fetch_json(URL) |> dd' 25c	needs LWP + JSON modules	needs net/http + json	needs urllib + json	curl -s URL | jq . ~40c
CSV → JSON	s 'csv_read("f") |> tj |> p' 28c	needs Text::CSV + JSON	needs csv + json	needs csv + json imports	—
Parallel map	s '1:1e6 |> pmap { $_ * 2 }' 29c	not built in	not built in	not built in	xargs -P8 50c+
Streaming parallel	s 'range(0,1e9) |> pmaps { $_ * 2 } |> take 10' 42c	not built in	not built in	not built in	not built in
Sparkline	s '(3,7,1,9) |> spark |> p' 27c	not built in	not built in	not built in	not built in
In-place sed (parallel)	s -i -pe 's/foo/bar/g' *.txt 28c	perl -i -pe 's/foo/bar/g' *.txt 33c (sequential)	ruby -i -pe '$_.gsub!(...)' 35c+	—	sed -i '' 's/foo/bar/g' *.txt 31c (sequential)

Feature matrix

Feature	stryke	perl5	ruby	python	awk	jq	nushell
No -e flag needed	yes	no	no	no (-c)	—	—	—
No semicolons	yes	no	yes	yes	yes	yes	yes
Built-in HTTP	yes	no	no	no	no	no	yes
Built-in JSON	yes	no	no	yes	no	yes	yes
Built-in CSV	yes	no	no	yes	no	@csv	yes
Built-in SQLite	yes	no	no	yes	no	no	yes
Parallel map/grep	yes	no	no	no	no	no	par-each
Pipe-forward |>	yes	no	no	no	no	|	|
Thread macro ~>	yes	no	no	no	no	no	no
In-place edit -i	parallel	sequential	sequential	no	no	no	no
Zsh glob qualifiers (/)/(.)/(L+N)/(om[1])	yes	no	no	no	no	no	no
Regex engine	3-tier	PCRE	Onigmo	re	ERE	PCRE	—
Data viz (spark/bars/flame)	yes	no	no	no	no	no	no
Clipboard (clip/paste)	yes	no	no	no	no	no	clip
$NR/$NF/$FNR AWK compat	yes	-MEnglish	no	no	native	no	no
Typed structs/enums/classes	yes	no	native	native	no	no	native
JIT compiler	Cranelift	no	YJIT	no	no	no	no
Single binary	44MB	system pkg	system pkg	system pkg	system pkg	3MB	50MB+

---

[0x02] USAGE

stryke 'p "Hello, world!"'                 # inline code — no -e needed
stryke 'p 1 + 2'                           # just quote and go
stryke script.stk arg1 arg2                  # script + args
stryke -lane 'p $F[0]'                     # bundled short switches
stryke -c script.stk                          # syntax check
stryke --lint script.stk                     # parse + compile (no run)
stryke --disasm script.stk                   # bytecode listing on stderr
stryke --ast script.stk                      # AST as JSON
stryke --fmt script.stk                      # pretty-print parsed source
stryke --profile script.stk                  # folded stacks + per-line/per-sub ns
stryke --flame script.stk                   # colored flamegraph bars in terminal
stryke --flame script.stk > flame.svg       # interactive SVG flamegraph when piped
stryke --explain E0001                      # expanded hint for an error code
stryke docs                                  # interactive reference book (vim-style: j/k/]/[/t/q)
stryke docs pmap                             # jump straight to a topic
stryke docs --toc                            # table of contents
stryke docs --search parallel                # search all pages
stryke 'doctor'                               # runtime health check — version / flags / paths / toolchain
stryke 'health'                               # alias for `doctor`
stryke 'lsp_words |> ep'                      # dump every name LSP tab-complete knows about
stryke 'p now'                                # current Unix-epoch seconds (alias of `time`)
stryke serve                                # static file server for $PWD on port 8000
stryke serve 8080 app.stk                   # HTTP server with handler script
stryke serve 3000 -e '"hello " . $req->{path}'  # one-liner HTTP server
stryke build script.stk -o myapp             # bake into a standalone binary ([0x0D])
stryke fmt -i .                              # format all .stk files recursively in place
stryke fmt lib/utils.stk                     # print formatted source to stdout
stryke minify app.stk                        # one-liner output (strip comments / POD, collapse newlines → `;`)
stryke minify -i lib/*.stk                   # minify in place — output still parses to the same AST
stryke check *.stk                           # parse + compile without executing (CI/editor)
stryke disasm script.stk                     # disassemble bytecode (learning/debugging)
stryke profile script.stk                    # run with profiling, structured output
stryke profile --flame script.stk -o out.svg # flamegraph to file
stryke bench                                 # run all benchmarks in bench/ or benches/
stryke init myapp                            # scaffold a new project (stryke.toml, lib/, t/, benches/)
stryke new myapp                             # alias for `init` that creates ./myapp/
stryke install                               # populate stryke.lock from stryke.toml (path deps; registry deps soon)
stryke add mylib --path=../mylib             # add a local path dep (registry deps land in RFC phase 7-8)
stryke remove mylib                          # drop a dep, regenerate stryke.lock
stryke tree                                  # print resolved dep graph from stryke.lock
stryke info mylib                            # show lockfile entry + store path for a dep
stryke repl                                  # start interactive REPL explicitly
stryke repl --load lib.stk                   # pre-load a library, then enter REPL
stryke lsp                                   # language server over stdio ([0x11])
stryke completions zsh                       # emit zsh completions to stdout
stryke ast script.stk                        # dump AST as JSON
stryke gen-docs                              # walk `.`, write Markdown docs for every .stk/.pl/.pm to docs/
stryke gen-docs lib --out site/api           # walk `lib/`, write to `site/api/` (mirrors source layout + index.md)
stryke prun *.stk                            # run multiple files in parallel
stryke -j 4 *.stk                             # run multiple files in parallel (4 threads)
stryke convert app.pl                        # convert Perl to stryke syntax with |> pipes
stryke deconvert app.stk                     # convert stryke back to Perl syntax
stryke --from-zsh script.zsh                 # transpile a zsh script to stryke (AST → AST)
stryke app.stk                                # warm starts skip parse + compile via ~/.stryke/scripts.rkyv ([0x0F])

-e is optional. If the first argument isn't a file on disk and looks like code, stryke runs it directly. stryke 'p 42' and stryke -e 'p 42' are equivalent. Use -e when combining with -n/-p/-l/-a (e.g. stryke -lane 'p $F[0]').

Semicolons

A newline ends a statement, so you do not need a trailing ; on each line. Use semicolons only when you put more than one statement on the same physical line.

val $answer = 40 + 2
p $answer                       # 42 — one statement per line, no `;` required

val $x = 1; val $y = 2; p $x + $y # 3 — same line needs `;` between statements

Interactive REPL

Run stryke with no arguments to enter a readline session: line editing, history (~/.stryke/history), tab completion for keywords, lexicals in scope, sub names, methods after -> on blessed objects, and file paths. exit/quit/Ctrl-D leaves. Non-TTY stdin is read as a complete program.

Edit mode (emacs / vi): the REPL defaults to emacs keybindings (Ctrl-A, Ctrl-E, etc.). To switch to vi modal editing (insert / normal modes, Esc to leave insert, h/j/k/l to navigate, etc.) drop this into ~/.stryke/config.toml:

[repl]
mode = "vi"   # or "emacs" (default)

STRYKE_REPL_MODE=vi stryke overrides the config for a single session — useful for sanity-checking either mode without editing the file.

__DATA__

A line whose trimmed text is exactly __DATA__ ends the program; the trailing bytes are exposed via the DATA filehandle.

Stdin / -n / -p / -i

echo data | stryke -ne 'print uc $_'        # line loop
cat f.txt | stryke -pe 's/foo/bar/g'        # auto-print like sed
stryke -i -pe 's/foo/bar/g' file1 file2     # in-place edit (parallel across files)
stryke -i.bak -pe 's/x/y/g' *.txt           # with backup suffix
echo a:b:c | stryke -aF: -ne 'print $F[1]'  # auto-split

-l chomps each record and sets $\. eof with no args is true on the last line of stdin or each @ARGV file (Perl-compat).

Text decoding — script reads, require, do, <>, backticks, par_lines, etc. all use UTF-8 when valid, else Latin-1 octets per line/chunk (matches stock perl tolerance). use open ':encoding(UTF-8)' switches <> to UTF-8 with U+FFFD replacement. slurp returns raw bytes (Perl-default byte-string semantics — length is byte count, spew round-trips exact bytes); text ops (eq, regex, substr) still work because byte values stringify via the same UTF-8-or-Latin-1 decoder on demand.

---

[0x03] PARALLEL PRIMITIVES

Each parallel block runs in its own interpreter context with captured lexical scope — no data races. Use varsync for shared counters. Optional progress => 1 enables an animated stderr bar (TTY) or per-item log lines (non-TTY).

# map / grep / sort / fold / for in parallel (list can be piped in)
# Three surface forms work for pmap/pgrep/pfor/pcache/pflat_map:
#   pmap { $_ * 2 } @list              # block form  ($_ = element)
#   pmap $_ * 2, @list                 # expression form
#   pmap double, @list                 # bare-fn form (sub double { $_0 * 2 })
val @doubled = @data |> pmap $_ * 2 , progress => 1
val @evens   = @data |> pgrep $_ % 2 == 0
val @sorted  = @data |> psort { _ <=> _1 }
val $sum     = @numbers |> preduce { _ + _1 }
pfor process, @items                  # pforeach is an alias, same as for/foreach
val @hashes  = pmap sha256, @blobs, progress => 1  # bare-fn

# ploop / pwhile — parallel loop / while: body runs as a worker loop on
# each of N rayon workers (default thread_count()); $_ = 1-based worker id.
# Desugar: pfor { while (COND) { BODY } } 1..N — last exits one worker.
ploop { val $job = next_job(); last unless defined $job; handle($job) }
ploop 4 { ... }                       # explicit worker count
varsync $run = 1
pwhile ($run) { $run = 0 if poll_shutdown(); step() }
pwhile 4 ($run) { ... }               # explicit worker count

# streaming parallel — lazy iterators, bounded memory, output as it completes
range(0, 1e9) |> pmaps { expensive(_) } |> take 10 |> ep  # stops after 10 results
range(0, 1e6) |> pgreps { is_prime(_) } |> ep              # parallel filter, streaming
range(0, 1e6) |> pflat_maps { [_, _ * 10] } |> ep         # parallel flat-map, streaming

# fused map+reduce, chunked map, memoized map, init fold
val $sum2     = @nums |> pmap_reduce { _ * 2 } { _ + _1 }
val @squared  = @million |> pmap_chunked 1000 { _ ** 2 }
val @once     = @inputs |> pcache expensive
val $hist     = @words |> preduce_init {}, { val ($acc, $x) = @_; $acc->{$x}++; $acc }

# fan — run a block or fn N times in parallel ($_/$_0 = index 0..N-1)
fan 8, work  # bare-fn form: fan N, FUNC
fan work, progress => 1  # uses rayon pool size (`stryke -j`)
fan 8 { work(_) }  # block form
fan { work(_) }  # block form, pool-sized
val @r = fan_cap 8, compute  # capture results in index order
val @r = fan_cap 8 { _ * _ }  # block form, capture

# pipelines — sequential or rayon-backed; same chain methods
val @r = (@data |> pipeline)->filter({ _ > 10 })->map({ _ * 2 })->take(100)->collect
### or 
val @r = @data |> pipeline |> filter $_ > 10 |> map $_ * 2 |> take 100 |> collect
val @r = @data |> par_pipeline |> filter  $_ > 10 |> map $_ * 2 |> collect

# multi-stage: batch (each stage drains list before next)
val $n = par_pipeline(
    source  => { readline(STDIN) },
    stages  => [ parse_json, transform ],
    workers => [4, 2],
    buffer  => 256,
)

# multi-stage: streaming (bounded crossbeam channels, concurrent stages, order NOT preserved)
val @r = ((1..1_000) |> par_pipeline_stream)->filter({ _ > 500 })->map({ _ * 2 })->collect()
## or
val @r = (1..1_000) |> par_pipeline_stream |> filter $_ > 500 |> map $_ * 2 |> collect

# channels + Go-style select
val ($tx, $rx) = pchannel(128)  # bounded; pchannel() is unbounded
val ($val, $idx) = pselect($rx1, $rx2)
val ($v, $i)     = pselect($rx1, $rx2, timeout => 0.5)  # $i == -1 on timeout

# barrier — N workers rendezvous
val $sync = barrier(3)
fan 3 { $sync->wait; p "all arrived" }

# persistent thread pool (avoids per-task spawn from pmap/pfor)
val $pool = ppool(4)
$pool->submit({ heavy_work(_) }) for @tasks
val @results = $pool->collect()

# parallel file IO
val @logs = "**/*.log" |> glob_par  # rayon recursive glob
par_lines "./big.log", { p if /ERROR/ }  # mmap + chunked line scan
par_walk  ".", { p if /\.rs$/ }  # parallel directory walk
par_sed qr/\bfoo\b/, "bar", @paths  # parallel in-place sed (returns # changed)
val @rs = par_find_files "src", "*.rs"  # parallel recursive file search by glob
val $n  = par_line_count @rs  # parallel line count across files

# native file watcher (notify crate: inotify/kqueue/FSEvents)
watch  "/tmp/x", p
pwatch "logs/*", heavy

# control thread count
stryke -j 8 -e '@data |> pmap heavy'

# distributed pmap over an SSH worker pool — see [0x10] for details
val $cluster = cluster(["build1:8", "build2:16"])
val @r = @huge |> pmap_on $cluster heavy

Parallel capture safety — workers set Scope::parallel_guard after restoring captured lexicals. Assignments to captured non-varsync aggregates are rejected at runtime; varsync, package-qualified names, and topics ($_/$a/$b) are allowed. pmap/pgrep treat block failures as undef/false; use pfor when failures must abort.

Nested implicit-param matrix _N<<<<< — world-first. Every block-form closure iter (grep { ... }, map { ... }, sort { ... }, ~> @arr map { ... }, fi { ... }, etc.) shifts an outer-topic chain across all positional slots, up to 5 frames back. Read the previous topic with _<, two back with _<<, up to five back with _<<<<<. Same for every positional slot: _1<<, _2<<<<<, etc. No other language has this — Clojure %, Scala _, Ruby _1, Swift $0, Raku $^a all stop at the current scope.

{} is the shift trigger. EXPR-form HOFs (grep EXPR, LIST, map EXPR, LIST, reject EXPR, LIST, grepv EXPR, LIST, |> grep EXPR, etc. — anything with no braces) do not shift the chain. The expression is evaluated in the surrounding lexical scope with $_/$_0 rebound per iter; everything else (including all positional aliases $_1, $_2, …) stays put. This makes higher-order combinator patterns work without chain-ascent boilerplate:

# Strain (Exercism): keep elements matching the predicate. The fn's args
# are `$_` (arrayref) and `$_1` (predicate coderef). EXPR-form `grep _1, …`
# evaluates `_1` per iter — and since there's no `{}` block, slot 1 stays
# bound to the caller's $_1 (the coderef), which the runtime then dispatches.
fn Exercism::Strain::keep    = [grep  _1, @$_]
fn Exercism::Strain::discard = [grepv _1, @$_]   # grepv ≡ reject (inverse)

The rule reads the same in both directions: {} → shift, no {} → no shift. Block boundaries are scope boundaries; expression positions are not.

Indexed-ascent shortcut _<N — past depth 2, counting chevrons gets error-prone. The lexer accepts _<N (where N is a positive integer) as syntactic sugar for _<<<...< (N chevrons). So _<3 ≡ _<<< (more readable past depth 2), _<5 ≡ _<<<<<, etc. Mixed forms work too: $_2<3 reaches positional 2 from 3 frames up. Disambiguator: _<3> and _<3:5> remain string-slice syntax; _<3 (without trailing > or :) is indexed-ascent.

The matrix:

slot 0 — bare `_` aliases `_0`, FOUR equivalent spellings per level:
  current  _    ≡ $_    ≡ _0    ≡ $_0
  1 up     _<   ≡ $_<   ≡ _0<   ≡ $_0<       (also: _<1, $_<1)
  2 up     _<<  ≡ $_<<  ≡ _0<<  ≡ $_0<<      (also: _<2, $_<2)
  3 up     _<<< ≡ $_<<< ≡ _0<<< ≡ $_0<<<     (also: _<3, $_<3)
  4 up     _<<<<  ≡ $_<<<<  ≡ _0<<<<  ≡ $_0<<<<   (also: _<4, $_<4)
  5 up     _<<<<< ≡ $_<<<<< ≡ _0<<<<< ≡ $_0<<<<<  (also: _<5, $_<5)

slot N ≥ 1 — two spellings per level (plus indexed form):
  current  _N   ≡ $_N
  1 up     _N<  ≡ $_N<        (also: _N<1, $_N<1)
  ...
  5 up     _N<<<<< ≡ $_N<<<<<  (also: _N<5, $_N<5)

The < glyph is iconic: "back/before/earlier" is universal in math and ASCII (<-, <<, version comparison). _ is "the topic" (Perl $_, Ruby _1, Scala _). Composition tells you the meaning at sight.

# Rolling difference — no temp var, no naming.
~> @prices map { _ - _< }
# Python: [prices[i]-prices[i-1] for i in range(1,len(prices))] — 41 chars
# Ruby:   prices.each_cons(2).map { |a,b| b-a }              — 38 chars
# stryke: ~> @prices map { _ - _< }                          — 24 chars

# 3-arg sub, reach back 4 closures from inside nested maps:
fn deep($_0, $_1, $_2) {
    ~> 1:1 map { ~> 1:1 map { ~> 1:1 map { ~> 1:1 map {
        # _N<<<< reads the Nth positional of `deep`
        _0<<<< . "," . _1<<<< . "," . _2<<<<      # "alpha,beta,gamma"
    } } } }
}
deep("alpha", "beta", "gamma")

# Cartesian-style sum across two arrays, golf form:
~> @outer pmap { ~> @inner pmap { _< + _ } } sum
# (`_<` rolls through previous topics across iter boundaries — same primitive
# powers running totals, moving averages, deltas)

# fan / fan_cap also rebind topic per worker:
$_ = 100
val @r = fan_cap 3 { $_< }                      # (100, 100, 100)
fan_cap 1 { $_ = "inner"; "$_< $_" }           # "outer inner"
$_ = 50; ~> 10 >{ $_ + $_< }                   # 60

Implementation: strykelang/scope.rs::set_closure_args shifts every active slot's chain on each frame entry; strykelang/lexer.rs lexes _<+, _N<+, and the indexed-ascent forms _<N/_M<N (bare and $-prefixed) as single tokens. Depth cap is hardcoded at 5 levels (debug_assert!(level <= 5) in scope.rs::topic_slot_key); past depth 5 the chain falls off and reads return undef. Bumping the cap is a one-line change.

Mutation semantics — topic variants align with |param| block params

A user writing $_ = ... or $_< = ... inside a block mutates only the current frame. Topic variants follow the exact same rule as |$x| block params and inner my $x: writes do not leak outward. The chain shift on the next frame entry remains purely a function of the outer topic value, never the inner mutation.

form	mutation propagates to outer scope?	mechanism
|$x| block param	NO — frame-local	param binding lives in callee frame
my $x inside a block	NO — frame-local	new lexical binding in current frame
my $x outer + inner closure writes $x	rejected at compile time	DESIGN-001 (closures capture by value)
varsync $x outer + inner closure writes $x	YES — explicit Arc<Mutex> opt-in	shared cell, atomic compound ops
our $x	YES — package-global by design	symbol table, not lexical
$_, $_<, $_<<, $_<<<, $_<<<<, $_<<<<<	NO — frame-local	Frame::set_scalar_raw bypasses CaptureCell write-through
$_0, $_1, … $_N and $_N<+ chain forms	NO — frame-local	same path as topic-chain writes

Implementation: strykelang/scope.rs::Scope::set_scalar recognizes topic-variant names via is_topic_variant_name (regex ^_[0-9]*<*$) and routes the write through Frame::set_scalar_raw, which bypasses the CaptureCell write-through that named outer-scope my variables use. Result: $_< always reads the lexical outer-scope topic of the current closure, never an in-flight mutation from a sibling iteration.

---

[0x04] SHARED STATE (varsync)

varsync declares variables backed by Arc<Mutex> shared across parallel blocks. Compound ops (++, +=, .=, |=, &=) hold the lock for the full read-modify-write cycle — fully atomic.

varsync $counter = 0
fan 10000 { $counter++ }  # always exactly 10000
p $counter

varsync @results
(1..100) |> pfor { push @results, $_ * $_ }

varsync %histogram
(0..999) |> pfor { $histogram{$_ % 10} += 1 }

# deque() and heap(...) already use Arc<Mutex<...>> internally
varsync $q  = deque()
varsync $pq = heap { _ <=> _1 }

For varsync scalars holding a Set, |/& are union/intersection. Without varsync, each thread gets an independent copy.

---

[0x05] NATIVE DATA SCRIPTING

Area	Builtins
HTTP (ureq)	fetch, fetch_json, fetch_async, await fetch_async_json, par_fetch, serve
JSON (serde_json)	json_encode, json_decode
CSV (csv)	csv_read (AoH), csv_write, par_csv_read
DataFrame	dataframe(path) → columnar; ->filter, ->group_by, ->sum, ->nrow, ->ncol
SQLite (rusqlite, bundled)	sqlite(path) → ->exec, ->query, ->last_insert_rowid
TOML / YAML	toml_decode, yaml_decode
Crypto	sha1, sha224, sha256, sha384, sha512, md5, hmac, hmac_sha256, crc32, uuid, base64_encode/decode, base64url_encode/decode, hex_encode/decode
Steganography (image, sha2, crc32fast) — world-first polymorphic stego builtin	hide(CARRIER, SECRET [, KEY]) / reveal(STEGO [, KEY]) / hide_capacity(CARRIER) — auto-dispatches PNG LSB (R/G/B, alpha skipped) for \x89PNG bytes vs zero-width-char text stego (U+200B/U+200C between visible chars) for everything else. 4-byte length + 4-byte CRC32 envelope catches tampering. Optional KEY enables SHA-256(key‖counter)-derived XOR mask on the secret.
Compression (flate2, zstd)	gzip, gunzip, zstd, zstd_decode
Time (chrono, chrono-tz)	datetime_utc, datetime_from_epoch, datetime_parse_rfc3339, datetime_strftime, datetime_now_tz, datetime_format_tz, datetime_parse_local, datetime_add_seconds, strptime, time_ago, elapsed
Structs / Enums / Classes / Types	struct Point { x => Float }, enum Color { Red, Green } (exhaustive match), class Dog extends Animal { breed: Str; fn bark { } }, abstract class/final class, trait Printable { fn to_str } (enforced, default method inheritance), pub/priv/prot visibility, static count: Int, BUILD/DESTROY, final fn, methods()/superclass()/does(), static::method(), typed val $x : Int, parametric containers List<Str> / Map<Int, Str>, return types fn add($x: Int): Int, --static mandatory typing ([0x08c])
Cyberpunk Terminal Art	cyber_city (neon cityscape), cyber_grid (synthwave perspective grid), cyber_rain/matrix_rain (digital rain), cyber_glitch/glitch_text (text corruption), cyber_banner/neon_banner (block-letter banners), cyber_circuit (circuit board), cyber_skull, cyber_eye — all output ANSI-colored Unicode art

val $data = "https://api.example.com/users/1" |> fetch_json
p $data->{name}

# Built-in HTTP server — one-liner web API
serve 8080, fn ($req) {
    # $req = { method, path, query, headers, body, peer }
    val $data = +{ path => $req->{path}, method => $req->{method} }
    status => 200, body => json_encode($data)
}
# or with workers: serve 8080, $handler, { workers => 16 }
# JSON content-type auto-detected; undef returns 404

val @rows = "data.csv" |> csv_read
val $df   = "data.csv" |> dataframe
val $db   = "app.db" |> sqlite
$db->exec("CREATE TABLE t (id INTEGER, name TEXT)")

# ─── Structs ────────────────────────────────────────────────────────
# Declaration: typed fields, optional defaults, or bare (Any type)
struct Point { x => Float, y => Float }
struct Point { x => Float = 0.0, y => Float = 0.0 }  # with defaults
struct Pair { key, value }  # untyped (Any)

# Construction: function-call, positional, or traditional ->new
val $p = Point(x => 1.5, y => 2.0)  # function-call with named args
val $p = Point(1.5, 2.0)  # positional (declaration order)
val $p = Point->new(x => 1.5, y => 2.0)  # traditional OO style
val $p = Point()  # uses defaults if defined

# Field access: getter (0 args) or setter (1 arg)
p $p->x  # 1.5 — getter
$p->x(3.0)  # setter
p $p->x  # 3.0

# User-defined methods
struct Circle {
    radius => Float,
    fn area { 3.14159 * $self->radius ** 2 }
    fn scale($factor: Float) {
        Circle(radius => $self->radius * $factor)
    }
}
val $c = Circle(radius => 5)
p $c->area  # 78.53975
p $c->scale(2)  # Circle(radius => 10)

# Built-in methods
val $q = $p->with(y => 5)  # functional update — new instance
val $h = $p->to_hash  # { x => 3.0, y => 5 }
val @f = $p->fields  # (x, y)
val $c = $p->clone  # deep copy

# Smart stringify — print shows struct name and fields
p $p  # Point(x => 3, y => 2)

# Structural equality — compares all fields
val $a = Point(1, 2)
val $b = Point(1, 2)
p $a == $b  # 1 (equal)
# ────────────────────────────────────────────────────────────────────

# ─── Enums (algebraic data types) ───────────────────────────────────
# Declaration: variants with optional typed data
enum Color { Red, Green, Blue }  # unit variants (no data)
enum Maybe { None, Some => Any }  # Some carries any value
enum Result { Ok => Int, Err => Str }  # typed data per variant

# Construction: Enum::Variant() syntax
val $c = Color::Red()  # unit variant
val $m = Maybe::Some(42)  # variant with data
val $r = Result::Err("not found")  # typed variant

# Smart stringify — shows enum name, variant, and data
p $c  # Color::Red
p $m  # Maybe::Some(42)
p $r  # Result::Err(not found)

# Type checking on variants with data
# Result::Ok("bad")  # ERROR: expected Int
# Maybe::Some()  # ERROR: requires data
# Color::Red(42)  # ERROR: does not take data

# Exhaustive enum matching — all variants must be covered or use `_` catch-all
val $light = Light::On()
val $s = match ($light) {
    Light::On()  => "on",
    Light::Off() => "off",
}
# Missing a variant without `_` → error:
# match ($c) { Color::Red() => "r" }  # ERROR: missing variant(s) Green, Blue
# ────────────────────────────────────────────────────────────────────

# ─── Cyberpunk Terminal Art ────────────────────────────────────────
p cyber_banner("STRYKE")          # large neon block-letter banner
p cyber_city()                    # procedural neon cityscape (80x24)
p cyber_city(120, 40, 99)         # custom width, height, seed
p cyber_grid(80, 20)              # synthwave perspective grid
p cyber_rain(80, 24)              # matrix-style digital rain
p cyber_glitch("BREACH", 7)       # glitch-distort text (intensity 1-10)
p cyber_circuit(60, 20)           # circuit board with traces and nodes
p cyber_skull()                   # neon skull (or "large" for big version)
p cyber_eye("large")              # all-seeing eye motif
# All output ANSI-colored Unicode — pipe to `less -R` or print directly.
# ────────────────────────────────────────────────────────────────────

# ─── Classes (full OOP) ────────────────────────────────────────────
# Declaration: class Name extends Parent impl Trait { fields; methods }
class Animal {
    name: Str
    age: Int = 0
    fn speak { p "Animal: " . $self->name }
}

# Inheritance with extends
class Dog extends Animal {
    breed: Str = "Mixed"
    fn bark { p "Woof! I am " . $self->name }
    fn speak { p $self->name . " barks!" }  # override
}

# Construction: named or positional
val $dog = Dog(name => "Rex", age => 5, breed => "Lab")
val $dog = Dog("Rex", 5, "Lab")  # positional

# Field access: getter (0 args) or setter (1 arg)
p $dog->name        # Rex
$dog->age(6)        # setter
p $dog->age         # 6

# Instance methods
$dog->bark()        # Woof! I am Rex
$dog->speak()       # Rex barks!

# Static methods: fn Self.name
class Math {
    fn Self.add($a, $b) { $a + $b }
    fn Self.pi { 3.14159 }
}
p Math::add(3, 4)   # 7
p Math::pi()        # 3.14159

# Traits (interfaces)
trait Printable { fn to_str }
class Item impl Printable {
    name: Str
    fn to_str { $self->name }
}

# Multiple inheritance
class C extends A, B { }

# isa checks inheritance chain
p $dog->isa("Dog")     # 1
p $dog->isa("Animal")  # 1
p $dog->isa("Cat")     # "" (false)

# Built-in methods (same as struct)
val @f = $dog->fields()       # (name, age, breed)
val $h = $dog->to_hash()      # { name => "Rex", ... }
val $d2 = $dog->with(age => 1) # functional update
val $d3 = $dog->clone()       # deep copy

# Smart stringify
p $dog  # Dog(name => Rex, age => 5, breed => Lab)

# Visibility (pub/priv/prot)
class Secret {
    pub visible: Int = 1
    priv hidden: Int = 42
    prot internal: Str = "base"         # subclass-only access
    pub fn get_hidden { $self->hidden } # internal access ok
}
class Child extends Secret {
    fn get_internal { $self->internal }  # prot: ok from subclass
}

# Abstract classes — cannot be instantiated; abstract methods enforced
abstract class Shape {
    name: Str
    fn area            # abstract method (no body) — subclasses must implement
    fn kind { "shape" } # concrete method — inherited by subclasses
}
class Circle extends Shape {
    radius: Float
    fn area { 3.14159 * $self->radius * $self->radius }
}
# Shape() → error!  Circle(name => "c", radius => 5) → ok
# class BadShape extends Shape { }  # → error: must implement abstract method `area`

# Static fields (class variables) — shared across all instances
class Counter {
    static count: Int = 0
    name: Str
    fn BUILD { Counter::count(Counter::count() + 1) }
}
val $a = Counter(name => "a")
val $b = Counter(name => "b")
p Counter::count()  # 2

# BUILD constructor hook — runs after field init, parent BUILD first
class Logger {
    log: Str = ""
    fn BUILD { $self->log("initialized") }
}

# DESTROY destructor — explicit via $obj->destroy(), child first
class Resource {
    fn DESTROY { p "cleanup" }
}
val $r = Resource()
$r->destroy()  # prints "cleanup"

# Trait enforcement — required methods checked at class definition
trait Drawable { fn draw }
# class Oops impl Drawable { }  # → error: missing required method `draw`
class Box impl Drawable {
    fn draw { "drawn" }    # satisfies trait contract
}
p Box()->does("Drawable")  # 1

# Trait default methods — inherited by implementing classes, overridable
trait Greetable {
    fn greeting { "Hello" }  # default method (has body)
    fn name                  # required method (no body)
}
class Person impl Greetable {
    n: Str
    fn name { $self->n }
    # greeting inherited from trait — Person()->greeting() returns "Hello"
}
class FormalPerson impl Greetable {
    n: Str
    fn name { $self->n }
    fn greeting { "Good day" }  # override the default
}

# Final classes — cannot be extended
final class Singleton { value: Int = 1 }
# class Bad extends Singleton { }  # → error

# Final methods — cannot be overridden
class Secure {
    final fn id { 42 }
    fn label { "secure" }  # can be overridden
}

# Reflection: methods(), superclass()
val @m = $dog->methods()     # ("speak", "bark", ...)
val @p = $dog->superclass()  # ("Animal")

# Late static binding: static::method() resolves to runtime class
class Base {
    fn class_name { static::identify() }
    fn identify { "Base" }
}
class Child extends Base {
    fn identify { "Child" }
}
Child()->class_name()  # "Child" (not "Base")

# Operator overloading for native classes
class Vec2 {
    x: Int; y: Int
    fn op_add($other) {
        Vec2(x => $self->x + $other->x, y => $self->y + $other->y)
    }
    fn op_eq($other) { $self->x == $other->x && $self->y == $other->y }
    fn stringify { "(" . $self->x . "," . $self->y . ")" }
}
val $v = Vec2(x => 1, y => 2) + Vec2(x => 3, y => 4)
p $v  # (4,6)
# Supported: op_add op_sub op_mul op_div op_mod op_pow op_concat
#            op_eq op_ne op_lt op_gt op_le op_ge op_spaceship op_cmp
#            op_neg op_bool op_abs op_numify stringify
# ────────────────────────────────────────────────────────────────────

typed val $n : Int = 42

# Typed fn parameters — runtime type checking on call
val $add = fn ($a: Int, $b: Int) { $a + $b }
p $add->(3, 4)  # 7
# $add->("x", 1)  # ERROR: sub parameter $a: expected Int

fn greet ($name: Str) { "Hello, $name!" }
p greet("world")  # Hello, world!

# stringify/str — convert any value to a parseable stryke literal
val $data = {a => [1, 2], b => "hello"}
val $s = str $data  # +{a => [1, 2], b => "hello"}
val $copy = eval $s  # round-trip via eval
p $copy->{a}[0]  # 1

# stringify works with functions (first-class serialization)
val $f = fn ($x: Int) { $x * 2 }
p str $f  # fn ($x: Int) { $x * 2; }
val $f2 = eval str $f  # round-trip: deserialize back to callable
p $f2->(21)  # 42

# streaming range — bidirectional lazy iterator
range(1, 5) |> e p                          # 1 2 3 4 5
range(5, 1) |> e p                          # 5 4 3 2 1

Sets

Native sets deduplicate by value (internal canonical keys; insertion order preserved for ->values). Use the set(LIST) builtin or Set->new(LIST); |> can supply the list. | / & are union / intersection when either side is a set (otherwise bitwise int ops).

val $s = set(1, 2, 2, 3)  # 3 members
val $t = (1, 1, 2, 4) |> set
val $u = $s | $t  # union
val $i = $s & $t  # intersection
$s->has(2)  # 1 / 0  (also ->contains / ->member)
$s->size  # count (->len / ->count)
val @v = $s->values  # array in insertion order

# varsync: compound |= and &= update shared sets (see [0x04])

---

[0x06] ASYNC / TRACE / TIMER

# async / spawn / await — lightweight structured concurrency
val $data = async { "https://example.com/" |> fetch }
val $file = spawn { "big.csv" |> \&slurp }
print await($data), await($file)

# trace varsync mutations to stderr (under fan, lines tagged with worker index)
varsync $counter = 0
trace { fan 10 { $counter++ } }

# timer / bench — wall-clock millis; bench returns "min/mean/p99"
val $ms     = timer heavy_work
val $report = bench heavy_work 1000

# eval_timeout — runs block on a worker thread; recv_timeout on main
eval_timeout 5 slow

# retry / rate_limit / every (tree interpreter only)
retry http_call times => 3, backoff => exponential
rate_limit(10, "1s") hit_api
every "500ms" tick

# generators — lazy `yield` values
val $g = gen { yield $_ for 1..5 }
val $next = $g->next  # [value, more]

---

[0x06b] AOP — BEFORE / AFTER / AROUND ADVICE

Aspect-oriented advice on user subs. Glob pointcuts, three advice kinds, proceed() for around. Same surface as zshrs's intercept builtin (zshrs/src/exec.rs), adapted to a real language: keyword statements instead of a CLI builtin.

# Before — runs before the matched sub. Sees $INTERCEPT_NAME, @INTERCEPT_ARGS.
before "fetch" { warn "calling fetch with @INTERCEPT_ARGS" }

# After — runs after. Sees $INTERCEPT_RESULT, $INTERCEPT_MS, $INTERCEPT_US.
after "fetch" { warn "fetch returned $INTERCEPT_RESULT in ${INTERCEPT_MS}ms" }

# Around — wraps. Must call proceed() to invoke the original.
around "expensive" {
    val $cached = cache_get($INTERCEPT_ARGS[0]);
    return $cached if defined $cached;
    val $r = proceed();
    cache_put($INTERCEPT_ARGS[0], $r);
    $r
}

# Glob patterns: *, ?
before "log_*"  { ... }     # any sub starting with log_
before "*"       { ... }     # every sub call

# Management
val @list = intercept_list();   # [[id, kind, pattern], ...]
intercept_remove($id);         # by id
intercept_clear();             # drop all

Semantics:

• Multiple before / after advices on the same name all fire (registration order).
• The first matching around wraps; later around matches on the same name are skipped (mirrors zshrs run_intercepts).
• around is AspectJ-style: the block's evaluated value is the call's return. proceed() runs the original and returns its value; the block can transform (proceed() + 100), forward (proceed()), or replace (omit the call and return a value directly).
• Recursion guard: calling the advised sub from inside its own advice runs the original directly without re-firing advice (no infinite loop).
• Coverage: user-defined subs only. Builtins (print, pmap, etc.) are not interceptable in v1.
• Pattern is a string literal ("foo", "log_*"); the leading keyword only commits to advice parsing when followed by a string literal, so before(...) as a normal call still works.
• Advice bodies are lowered to bytecode at compile time and dispatched through the VM (run_block_region) — the same path used by map { } / grep { } blocks. This keeps compile-time name resolution (our-qualified scalars, lexical slots) consistent inside advice and outside it. The tree-walker is banned from the advice path; see tests/tree_walker_absent_aop.rs for the source-level invariant.
• Body lowering requires the final statement to be an expression (the same constraint as map { } block lowering). Bodies that end in a literal for/while/if block, or contain a literal return, are rejected at advice-firing time with a runtime error — rewrite the body so it ends in an expression and avoids early-return.

Builtins from inside advice bodies:

• proceed() — only legal inside around; runs the original sub with the saved args, returns its value.
• intercept_list() — returns [[id, kind, pattern], ...] for all registered advices.
• intercept_remove($id) — removes one by id; returns the count removed (0 or 1).
• intercept_clear() — drops all; returns count cleared.

---

[0x07] CLI FLAGS

All stock perl flags are supported: -0, -a, -c, -C, -d, -D, -e, -E, -f, -F, -g, -h, -i, -I, -l, -m, -M, -n, -p, -s, -S, -t, -T, -u, -U, -v, -V, -w, -W, -x, -X. Perl-style single-dash (-version, -help) and GNU-style double-dash (--version, --help) long forms work. Bundled switches are expanded: -Mstrict → -M strict, -I/tmp → -I /tmp, -V:version → -V version, -lane → -l -a -n -e.

stryke-specific long flags:

Flag	Description
--lint / --check	Parse + compile bytecode without running
--disasm / --disassemble	Print bytecode disassembly to stderr before VM execution
--ast	Dump parsed AST as JSON and exit
--fmt	Pretty-print parsed Perl to stdout and exit
--from-zsh	Transpile a zsh script to idiomatic stryke on stdout and exit (in-process zsh AST → stryke AST; externals → system, builtins → native, val/var declarations)
--profile	Wall-clock profile: per-line + per-sub timings on stderr
--flame	Flamegraph: colored terminal bars when interactive, SVG when piped (stryke --flame x.stk > flame.svg)
--record	Record one row per stryke run (wall-clock, exit code, argv) to ~/.stryke/perf.sqlite. Inherits to child processes via STRYKE_RECORD=1 env, so s --record t TESTS... records one row per test file. Query via the perfview builtin.
--no-jit	Disable Cranelift JIT (bytecode interpreter only)
--compat	Perl 5 strict-compatibility mode: disable all stryke extensions (|>, struct, enum, match, pmap, #{expr}, etc.)
--static	Mandatory static typing (Kotlin-style): every function parameter, return type, and variable declaration must be typed or inferable; statically-known type mismatches abort before the program runs. See [0x08c]
--no-interop	Reject Perl-isms (sub, say, reverse, scalar, $a/$b outside sort blocks); force idiomatic stryke (fn, p, rev, len, $_0/$_1). See [0x08a]
--explain CODE	Print expanded hint for an error code (e.g. E0001)
--lsp	Language server over stdio ([0x11])
--dap [HOST:PORT]	Debug Adapter Protocol server. With no arg → stdio; with HOST:PORT → connects to a TCP socket the spawner is listening on. Used by the JetBrains plugin under editors/intellij/
-d	TTY debugger (perl -d style REPL on stdin/stderr). Use with --: stryke -d -- script.stk
-j N / --threads N	Set number of parallel threads (rayon)
--remote-worker	Persistent cluster worker over stdio ([0x10])
--remote-worker-v1	Legacy one-shot cluster worker over stdio
build SCRIPT [-o OUT]	AOT compile script to standalone binary ([0x0D])
doc [TOPIC]	Interactive reference book with vim-style navigation (stryke doc, stryke doc pmap, stryke doc --toc)
serve [PORT] [SCRIPT]	HTTP server (default port 8000): static files (stryke serve), script (stryke serve 8080 app.stk), one-liner (stryke serve 3000 -e 'EXPR')
fmt [-i] FILE...	Format source files in place or to stdout (stryke fmt -i . formats all recursively)
minify [-i] FILE...	Strip comments / POD / blank lines and collapse statements onto a single line with ; separators. Output still parses to the same AST as the input.
check FILE...	Parse + compile without executing; report errors with file:line:col (CI/editor integration)
disasm FILE	Disassemble bytecode to stderr (learning the VM, debugging perf)
profile [--flame] [--json] FILE	Run with profiling; --flame generates SVG, -o FILE writes to file
bench [FILE|DIR]	Discover and run benchmarks from bench/ or benches/ (bench_*.stk, b_*.stk)
init [NAME]	Scaffold a new project: main.stk, lib/, bench/, t/, .gitignore
repl [--load FILE]	Start interactive REPL explicitly, with optional pre-loaded file
lsp	Start Language Server Protocol over stdio (equivalent to --lsp)
completions [SHELL]	Emit shell completions to stdout (stryke completions zsh > _stryke)
ast FILE	Dump parsed AST as JSON to stdout
prun FILE...	Run multiple script files in parallel using all cores
convert [-i] FILE...	Convert Perl source to stryke syntax with |> pipes
deconvert [-i] FILE...	Convert stryke .stk files back to standard Perl syntax

getopts builtin — Getopt::Long-style argv parsing

For parsing your own script's argv (@ARGV), stryke ships a getopts builtin shaped after Perl's Getopt::Long. It returns a hash of canonical-name → value, and removes consumed options from @ARGV so the leftover positional arguments stay behind.

When called with just SPECS, getopts operates on @ARGV directly — no \@ARGV boilerplate needed:

val %opts = %{ getopts([
    "verbose|v",         # bool flag (present = 1)
    "file|f=s",          # required string
    "count|n=i",         # required int
    "rate=f",            # required float
    "out:s",             # optional string ("" if absent)
    "tag|t=s@",          # repeatable → arrayref
    "define|D=s%",       # --define key=val → hashref
    "debug+",            # incremental: -ddd → 3
    "color!",            # negatable: --no-color → 0
]) };

# Hash form lets each spec carry a default:
val %opts = %{ getopts({
    "verbose|v" => 0,
    "count|n=i" => 10,
    "tag|t=s@"  => [],
}) };

# Explicit-array-ref form (parse a list other than @ARGV):
val @args = ("--verbose", "file.txt");
val %opts = %{ getopts(\@args, [ "verbose|v" ]) };

The first argument is interpreted as follows: an array ref of spec strings or a hash ref of specs → operates on @ARGV; an array ref followed by a SPECS argument → explicit-argv form. The two-arg (SPECS, META) form is recognised when the second argument is a hash ref containing only prog/desc/epilog keys (see auto-help below).

Spec language (subset of Perl's Getopt::Long):

Spec	Meaning
name	Bool flag, no argument
n|name	Same option, multiple names; first is canonical
name=s	Required string arg
name=i	Required int arg
name=f	Required float arg
name:s	Optional string (defaults to "" if flag is given without)
name:i	Optional int (defaults to 0)
name:f	Optional float (defaults to 0.0)
name=s@	Repeatable → arrayref (=i@ / =f@ typed)
name=s%	--name key=val → hashref (=i% / =f% typed)
name!	Negatable bool; --no-name sets 0
name+	Counter: each occurrence increments by 1

Parsing rules:

• --name, --name=value, and --name value all accepted for long options.
• -n, -n value, and -nvalue all accepted for short options.
• Bundling: -vDR is parsed as -v -D -R. If a char in the bundle takes an argument, the rest of the token becomes its value (-vfx.txt → -v -f x.txt). Bundling and = are mutually exclusive — -X=value is treated as a single short option name with an inline value.
• -- terminates option parsing; everything after is positional.
• Numeric tokens (-5, -3.14) are treated as positionals.
• The first non-option positional stops parsing (no intermixed mode in v1).
• Unknown options or type mismatches raise a runtime error.

Defaults when an option is absent: bool / negatable bool / counter → 0; =s@ → empty arrayref; =s% → empty hashref; required scalars (=s/=i/=f) → not present in the returned hash unless given a default via the hash form.

Per-option metadata + auto---help

The hash form also accepts a hashref value carrying per-option metadata (help, default, required, metavar). When any spec has a help string and the user hasn't claimed their own --help/-h, getopts intercepts --help/-h in the input, prints a formatted usage block to stdout, and exit(0)s.

val %opts = %{ getopts({
    "verbose|v" => { help => "enable verbose output" },
    "file|f=s"  => { help => "output path", default => "out.txt" },
    "count|n=i" => { help => "iterations", required => 1 },
    "tag|t=s@"  => { help => "tag (repeatable)" },
    "color!"    => { help => "colored output" },
}, {
    prog   => "myscript",        # banner program name (default: argv[0] basename)
    desc   => "do a thing",      # banner description line
    epilog => "see docs",        # trailing line after the option list
}) };

Metadata keys (D1 form):

Key	Meaning
help	Help-text shown next to the option in --help output
default	Default value if the option is not given
required	Error at end of parse if the option (or its default) is missing
metavar	Override the placeholder shown next to the option in help output

Scalar and hashref values can be mixed in the same spec hash — a bare scalar value is still treated as the default:

val %opts = %{ getopts({
    "verbose|v" => 0,                                # default only (legacy form)
    "file|f=s"  => { help => "output path" },        # D1 metadata
}) };

--help output for the first example above:

Usage: myscript [OPTIONS]

do a thing

Options:
  -v, --verbose        enable verbose output
  -f, --file VALUE     output path (default: out.txt)
  -n, --count N        iterations (required)
  -t, --tag VALUE      tag (repeatable)
  --color, --no-color  colored output
  -h, --help           show this help and exit

see docs

---

[0x08] SUPPORTED PERL FEATURES

Data

Scalars $x, arrays @a, hashes %h, refs \$x/\@a/\%h/\&sub, anon [...]/{...}, code refs / closures (capture enclosing lexicals), qr// regex objects, blessed references, native sets (set(LIST) / Set->new(...)), deque(), heap().

Control flow

if/elsif/else/unless, while/until, do { } while/until, C-style for, foreach, last/next/redo with labels, postfix if/unless/while/until/for, ternary, try { } catch ($err) { } finally { }, given/when/default, algebraic match (EXPR) { PATTERN [if EXPR] => EXPR, ... } (regex, array, hash, wildcard, literal patterns; bindings scoped per arm; exhaustive enum variant checking), eval_timeout SECS { ... }.

Operators

Arithmetic, string ./x, comparison (including Raku-style chained comparisons like 1 < $x < 10), eq/ne/lt/gt/cmp, logical &&/||////!/and/or/not (?? / ??= are stryke aliases for // / //=; null aliases undef), bitwise (|/& are set ops on native Set), assignment + compound (+=, .=, //=, …), regex =~/!~, range .. / ... (incl. flip-flop with eof), arrow ->, pipe-forward |> (stryke extension — threads the LHS as the first argument of the RHS call; see Extensions beyond stock Perl 5).

Regex engine

Three-tier compile (Rust regex → fancy-regex → PCRE2). Perl $ end anchor (no /m) is rewritten to (?:\n?\z). Match =~, dynamic $str =~ $pat, substitution s///, transliteration tr///, flags g/i/m/s/x/e/r, captures $1…$n, named groups → %+/$+{name}, \Q...\E, quotemeta, m///qr//. The /r flag (non-destructive) returns the modified string instead of the match count — auto-injected when s/// or tr/// appear as pipe-forward RHS. Bare /pat/ in statement/boolean context is $_ =~ /pat/.

Subroutines

fn name { } with optional prototype, typed parameters (fn add($a: Int, $b: Int)), default parameter values (fn greet($name = "world")), anon subs/closures, implicit return of last expression (VM), @_/shift/return, postfix return ... if COND, AUTOLOAD with $AUTOLOAD set to the FQN.

Built-ins (selected)

Category	Functions
Array	push, pop, shift, unshift, splice, splice_last (last removed — --no-interop replacement for scalar splice), rev (string / list reverse), sort, map, grep, filter, reduce, fold, fore, e, preduce, len/cnt/count (element count — replaces scalar @a), partition, min_by, max_by, zip_with, interleave, frequencies, tally, count_by, pluck, grep_v, head, tail, first
Hash	keys, values, each, delete, exists, select_keys, top, deep_clone/dclone, deep_merge/dmerge, deep_equal/deq
Functional	compose/comp, partial, curry, memoize/memo, once, constantly, complement, juxt, fnil
String	chomp, chop, length, substr, index, rindex, split, join, sprintf, printf, uc/lc/ucfirst/lcfirst, chr, ord, hex, oct, crypt, fc, pos, study, quotemeta, trim, lines, words, chars, digits, numbers, graphemes, columns, sentences, paragraphs, sections, snake_case, camel_case, kebab_case, template, shell_quote, shell_split
Binary	pack, unpack (subset A a N n V v C Q q Z H x w i I l L s S f d + *), unpack_first / unpack1 / up1 (first decoded element — --no-interop replacement for scalar unpack), vec
Numeric	abs, int, sqrt, squared/sq, cubed/cb, expt(B,E), sin, cos, atan2, exp, log, rand, srand, avg, stddev, clamp, normalize, range(N, M) (lazy bidirectional)
I/O	print, p, printf, open (incl. open my $fh, files, -| / |- pipes), close, eof, readline, read, seek, tell, sysopen, sysread/syswrite/sysseek, handle methods ->print/->p/->printf/->getline/->close/->eof/->getc/->flush, slurp, swallow/swa (glob → hash {abspath => bytes}), ingest/ing (streaming [path, bytes] iterator), burp (inverse of swallow — hash → files, mkdir -p), input, backticks/qx{}, capture (structured: ->stdout/->stderr/->exit), pager/pg/less (pipes value into $PAGER; TTY-gated), binmode, fileno, flock, getc, select, truncate, formline, read_lines, append_file, to_file, read_json, write_json, tempfile, tempdir, xopen/xo (system open — open on macOS, xdg-open on Linux), clip/clipboard/pbcopy (copy to clipboard), paste/pbpaste (read clipboard)
Directory	opendir, readdir, closedir, rewinddir, telldir, seekdir, files, filesf/f, fr (recursive files, lazy iterator), dirs/d, dr (recursive dirs, lazy iterator), sym_links, sockets, pipes, block_devices, char_devices
File tests	-e, -f, -d, -l, -r, -w, -s, -z, -x, -t (defaults to $_)
System	system, exec, exit, chdir, mkdir, unlink, rename, chmod, chown, chroot, stat, lstat, link, symlink, readlink, glob, glob_par, glob_match, which_all, par_sed, par_find_files, par_line_count, ppool, barrier, fork, wait, waitpid, kill, alarm, sleep, times, dump, reset
System Stats	mem_total, mem_free, mem_used, swap_total, swap_free, swap_used, disk_total, disk_free, disk_avail, disk_used, load_avg, sys_uptime, page_size, os_version, os_family, endianness, pointer_width, proc_mem/rss
Sockets	socket, bind, listen, accept, connect, send, recv, shutdown, socketpair
Network	gethostbyname, gethostbyaddr, getpwnam, getpwuid, getpwent/setpwent/endpwent, getgrnam, getgrgid, getgrent/setgrent/endgrent, getprotobyname, getprotobynumber, getservbyname, getservbyport
SysV IPC	msgctl, msgget, msgsnd, msgrcv, semctl, semget, semop, shmctl, shmget, shmread, shmwrite (stubs — runtime error)
Type	defined, undef, ref, bless, tied, untie, type_of, byte_size
Serialization	to_json, to_csv, to_toml, to_yaml, to_xml, to_html, to_markdown, to_table/tbl, to_jsonl/from_jsonl, to_ini/from_ini, ddump, stringify/str, json_encode/json_decode
Visualization	sparkline/spark, bar_chart/bars, flame/flamechart, histo, gauge, spinner, spinner_start/spinner_stop
Control	die, warn, eval, do, require, caller, wantarray, goto LABEL, continue { } on loops, prototype
Number Theory	prime_factors, divisors, num_divisors, sum_divisors, is_perfect, is_abundant, is_deficient, collatz_length, collatz_sequence, lucas, tribonacci, nth_prime, primes_up_to/sieve, next_prime, prev_prime, triangular_number, pentagonal_number, is_pentagonal, perfect_numbers, twin_primes, goldbach, prime_pi, totient_sum, subfactorial, bell_number, partition_number, multinomial, is_smith, aliquot_sum, abundant_numbers, deficient_numbers
Statistics	skewness, kurtosis, linear_regression, moving_average, exponential_moving_average, coeff_of_variation, standard_error, normalize_array, cross_entropy, euclidean_distance, minkowski_distance, mean_absolute_error, mean_squared_error, median_absolute_deviation, winsorize, weighted_mean
Geometry	area_circle, area_triangle, area_rectangle, area_trapezoid, area_ellipse, circumference, perimeter_rectangle, perimeter_triangle, polygon_area, sphere_volume, sphere_surface, cylinder_volume, cone_volume, heron_area, point_distance, midpoint, slope, triangle_hypotenuse, degrees_to_compass
Financial	npv, depreciation_linear, depreciation_double, cagr, roi, break_even, markup, margin, discount, tax, tip
Encoding	morse_encode/morse, morse_decode, nato_phonetic, int_to_roman, roman_to_int, binary_to_gray, gray_to_binary, pig_latin, atbash, braille_encode, phonetic_digit, to_emoji_num
Color	hsl_to_rgb, rgb_to_hsl, hsv_to_rgb, rgb_to_hsv, color_blend, color_lighten, color_darken, color_complement, color_invert, color_grayscale, random_color, ansi_256, ansi_truecolor, color_distance
Constants	pi, tau, phi, epsilon, speed_of_light, gravitational_constant, planck_constant, avogadro_number, boltzmann_constant, elementary_charge, electron_mass, proton_mass, i64_max, i64_min, f64_max, f64_min
Matrix	matrix_transpose, matrix_inverse, matrix_hadamard, matrix_power, matrix_flatten, matrix_from_rows, matrix_map, matrix_sum, matrix_max, matrix_min
DSP / Signal	convolution, autocorrelation, fft_magnitude, zero_crossings, peak_detect
Algorithms	next_permutation, is_balanced_parens, eval_rpn, merge_sorted, binary_insert, reservoir_sample, run_length_encode_str, run_length_decode_str, range_expand, range_compress, group_consecutive_by, histogram, bucket, clamp_array, normalize_range
Validation	luhn_check, is_valid_hex_color, is_valid_cidr, is_valid_mime, is_valid_cron, is_valid_latitude, is_valid_longitude
Text	ngrams, bigrams, trigrams, char_frequencies, is_anagram, is_pangram, mask_string, chunk_string, camel_to_snake, snake_to_camel, collapse_whitespace, remove_vowels, remove_consonants, strip_html, metaphone, double_metaphone, initials, acronym, superscript, subscript, leetspeak, zalgo, sort_words, unique_words, word_frequencies, string_distance, string_multiply, deburr
Misc	fizzbuzz, roman_numeral_list, look_and_say, gray_code_sequence, sierpinski, mandelbrot_char, game_of_life_step, tower_of_hanoi, pascals_triangle, truth_table, base_convert, roman_add, haversine, bearing, bmi, bac_estimate

Perl-compat highlights

• OOP — @ISA (incl. our @ISA outside main), C3 MRO (live, not cached), $obj->SUPER::method. tie for scalars/arrays/hashes with TIESCALAR/TIEARRAY/TIEHASH, FETCH/STORE, plus EXISTS/DELETE on tied hashes. tied returns the underlying object; untie removes the tie.

• use overload — 'op' => 'method' or \&handler; binary dispatch with (invocant, other), nomethod, unary neg/bool/abs, "" for stringification, fallback => 1.

• $? / $| — packed POSIX status from system/backticks/pipe close; autoflush on print/printf.

• $. — undef until first successful read, then last-read line count.

• AWK record vars (-n/-p loop) — $NR is the cumulative record number across all input files; $FNR is the per-file record number (resets at each file, like $.); $NF is the field count under -a autosplit. Available without -MEnglish; disabled under --compat.

• print/p/printf with no args — uses $_ (and printf's format defaults to $_).

• Bareword statement — name; calls a scwub with @_ = ($_).

• Typeglobs — *foo = \&bar, *lhs = *rhs copies sub/scalar/array/hash/IO slots; package-qualified *Pkg::name supported.

• %SIG (Unix) — SIGINT/SIGTERM/SIGALRM/SIGCHLD as code refs; handlers run between statements/opcodes via perl_signal::poll. IGNORE and DEFAULT honored.

• format / write — partial: format NAME = ... . registers a template; pictures @<<<<, @>>>>, @||||, @####, @****, literal @@. formline populates $^A. write (no args) uses $~ to stdout. Not yet: write FILEHANDLE, $^.

• @INC / %INC / require / use / import — @INC is built from -I, vendor/perl, system perl's @INC (set STRYKE_NO_PERL_INC to skip), the script dir, STRYKE_INC, then .. List utilities (sum, min, max, uniq, reduce, pairs, zip, mesh, …) are stryke-native bare-name builtins implemented in Rust at strykelang/list_builtins.rs — no Perl module shim, no module to import. use Module qw(a b); honors @EXPORT_OK/@EXPORT for actual user modules. import Module qw(a b); is a 1:1 alias for use — same parse, same AST, same semantics; spell it whichever way reads better for the file. Built-in pragmas (strict, warnings, utf8, feature, open, Env) do not load files. use Module VERSION (Perl 5's per-module pin syntax — use Foo::Bar 2.13;) routes resolution directly to ~/.stryke/store/<name>@<VERSION>/ — skips lockfile and store-scan; surfaces a hard error if that exact extraction isn't on disk (run s install). Plain use Module resolves three ways: (1) project-local lib/Foo.stk always wins; (2) inside a project, the stryke.lock pin is authoritative — missing store dir errors instead of falling through to a different version; (3) outside a project, the resolver scans ~/.stryke/store/<canonical>@*/ and picks the HIGHEST semver (2.0 > 1.99, 0.10.0 > 0.3.0; release beats pre-release so 1.0.0 > 1.0.0-rc1; within pre-releases rc10 > rc2 > rc1). The store-scan honors the namespace bridge (use GUI finds stryke-gui@*) and is independent of installed.toml — multiple versions of the same package coexist in the store and the newest wins unless you pin.

• chunked / windowed / fold — Use pipe-forward: LIST |> chunked(N), LIST |> windowed(N), LIST |> fold { BLOCK } (same for reduce). fold is an alias for reduce. List context → arrayrefs per chunk/window or the folded value; scalar context → chunk/window count where applicable.

  val @pairs = (1, 2, 3, 4) |> chunked(2)  # ([1,2], [3,4])
  val @slide = (1, 2, 3) |> windowed(2)  # ([1,2], [2,3])
  val @pipe  = (10, 20, 30) |> chunked(2)  # ([10,20], [30])
  val $sum   = (1, 2, 3, 4) |> fold { $a + $b }  # same as reduce
  val $cat   = qw(a b c) |> fold { $a . $b }
  

• use strict — refs/subs/vars modes (per-mode use strict 'refs' etc.). strict refs rejects symbolic derefs at runtime; strict vars requires a visible binding.

• BEGIN / UNITCHECK / CHECK / INIT / END — Perl order; ${^GLOBAL_PHASE} matches Perl.

• String interpolation — $var #{23 * 52}, $h{k}, $a[i], @a, @a[slice] (joined with $"), $#a in slice indices, $0, $1..$n. Escapes: \n \r \t \a \b \f \e \0, \x{hex}, \xHH, \u{hex}, \o{oct}, \NNN (octal), \cX (control), \N{U+hex}, \N{UNICODE NAME}, \U..\E, \L..\E, \u, \l, \Q..\E.

• zsh parameter expansion (world-first outside zsh) — double-quoted strings support the full zsh ${ … } grammar: defaults (${x:-d} ${x:=d} ${x:?m} ${x:+a}), trim (${x#p} ${x##p} ${x%p} ${x%%p}), substitution (${x/p/r} ${x//p/r}), length (${#x}), flags (${(U)x} ${(L)x} …), and subscripts + subscript flags (${a[2]} ${a[2,4]} ${a[(i)v]} ${a[(r)pat]}). Delegates to zshrs's paramsubst engine — single source of truth, zero reimplementation. Gated off under --compat, where ${ … } keeps pure Perl 5 semantics. For the statically-ambiguous forms (${#x} length-vs-count, positional defaults) call the zpexpand builtin with explicit-typed args, arrays by reference: zpexpand('${#arr}', "arr", \@arr) → element count.

• Triple-quoted strings — """...""" for interpolating multiline strings (same $var/@arr/#{expr}/escape rules as "..."); raw newlines preserved verbatim, no indent stripping. r"""...""" is the raw form: zero interpolation, zero backslash escapes — every byte copied literally until the closing """. "" (two quotes) inside the body does NOT close — only """ does.

• __FILE__ / __LINE__ — compile-time literals.

• Heredocs <<EOF, POD skipping, shebang handling, qw()/q()/qq() with paired delimiters.

• Special variables — large set of ${^NAME} scalars pre-seeded; see SPECIAL_VARIABLES.md. Still missing vs Perl 5: English, full $^V as a version object.

Extensions beyond stock Perl 5

• Native CSV (csv_read/csv_write), columnar dataframe, embedded sqlite.

• HTTP (fetch/fetch_json/fetch_async/par_fetch), JSON (json_encode/json_decode).

• Crypto, compression, time, TOML, YAML helpers (see [0x05]).

• All parallel primitives in [0x03] (pmap, fan, pipeline, par_pipeline_stream, pchannel, pselect, barrier, ppool, glob_par, par_walk, par_lines, par_sed, par_find_files, par_line_count, pwatch, watch).

• Distributed compute ([0x10]): cluster([...]) builds an SSH worker pool; pmap_on $cluster { } @list and pflat_map_on $cluster { } @list fan a map across persistent remote workers with fault tolerance and per-job retries.

• Standalone binaries ([0x0D]): stryke build SCRIPT -o OUT bakes a script into a self-contained executable.

• Inline Rust FFI ([0x0E]): rust { pub extern "C" fn ... } blocks compile to a cdylib on first run, dlopen + register as Perl-callable subs.

• Bytecode cache ([0x0F]): single rkyv shard at ~/.stryke/scripts.rkyv — mmap + zero-copy ArchivedHashMap lookup skips lex/parse/compile on warm starts. Disable with STRYKE_CACHE=0.

• Language server ([0x11]): stryke lsp runs an LSP server over stdio with diagnostics, hover, completion.

• varsync shared state ([0x04]).

• frozen my (or const my — same thing, more familiar spelling), typed my, struct, enum, class (full OOP with extends/impl), trait, algebraic match, try/catch/finally, eval_timeout, retry, rate_limit, every, gen { ... yield }.

• Raku-style chained comparisons — 1 < $x < 10 desugars to (1 < $x) && ($x < 10) at parse time. Works with all comparison operators (<, <=, >, >=, lt, le, gt, ge) and chains of any length.

• Default parameter values — fn greet($name = "world"), fn range(@vals = (1,2,3)), fn config(%opts = (debug => 0)). Defaults evaluated at call time when argument not provided.

• Functional composition — compose, partial, curry, memoize, once, constantly, complement, juxt, fnil:

  val $f = compose(fn { $_ + 1 }, fn { $_ * 2 })
  $f(5)  # 11 (double then inc)
  
  val $add5 = partial(fn { $_[0] + $_[1] }, 5)
  $add5(3)  # 8
  
  val $cadd = curry(fn { $_[0] + $_[1] }, 2)
  $cadd(1)(2)  # 3
  
  val $fib = memoize(fn { ... })  # cached by args
  val $init = once(fn { expensive_setup() })  # called at most once
  

• Deep structure utilities — deep_clone/dclone, deep_merge/dmerge, deep_equal/deq, tally:

  val $b = deep_clone($a)  # recursive deep copy
  val $m = deep_merge(\%a, \%b)  # recursive hash merge
  deep_equal([1,2,{x=>3}], [1,2,{x=>3}])  # 1 (structural eq)
  val $t = tally("a","b","a")  # {a => 2, b => 1}
  

• Bare _ as topic shorthand — in any expression position, bare _ is equivalent to $_. Inspired by Raku's WhateverCode and Scala's placeholder syntax. Enables ultra-concise blocks: map{_*2} instead of map{$_ * 2}. The sigil-free form compresses better — no spaces needed around _ when adjacent to operators.

• Outer topic $_< — access the enclosing scope's $_ from nested blocks; up to 5 levels ($_< through $_<<<<<, or the indexed form $_<5). See [0x03].

• fore (e) — side-effect-only list iterator (like map but void, returns item count). Works with { BLOCK } LIST, blockless e EXPR, LIST, and pipe-forward |> e p. Use for print/log/accumulator loops.

• Pipe-forward |> — parse-time desugaring (zero runtime cost); threads the LHS as the first argument of the RHS call, left-associative. map, grep/filter, sort, and e accept blockless expressions on the RHS of |> — no { } required for simple transforms:

  # chain HTTP fetch → JSON decode → jq filter
  val @titles = $url |> fetch_json |> json_decode |> json_jq '.articles[].title'
  
  # blockless list pipelines — no braces needed for simple expressions
  files |> filter /[a-e]/ |> e -f $_ && system("cat $_")
  "a".."z" |> map uc |> e p                      # A B C … Z
  "a".."z" |> grep /[aeiou]/ |> e p              # a e i o u
  "a".."z" |> filter 't' |> e p                  # t  (literal = equality test)
  1..10 |> filter $_ > 5 |> sort |> e p      # blocks still work
  1..5 |> map $_ * $_ |> join "," |> p  # 1,4,9,16,25
  
  # e — side-effect-only iteration (like map but void, returns count)
  qw(apple banana cherry) |> grep /^a/ |> map uc |> e p  # APPLE
  
  # unary builtins — `x |> length`, `x |> uc`, `x |> sqrt`, etc.
  "hello" |> length |> p  # 5
  16 |> sqrt |> p  # 4
  "ff" |> hex |> p  # 255
  
  # bareword on RHS becomes a unary call: `x |> f` → `f(x)`
  # call on RHS prepends: `x |> f(a, b)` → `f(x, a, b)`
  # map/grep/filter/sort/join/reduce/fold/e — LHS fills the list slot
  # chunked/windowed — `LIST |> chunked(N)` prepends the list before the size
  # scalar on RHS: `x |> $cr` → `$cr->(x)`
  
  # regex ops in pipelines — s///, tr///, and m// work as RHS of |>
  # s/// and tr/// auto-inject /r so the modified string flows through:
  "hello world" |> s/world/perl/  |> p  # hello perl
  "hello world" |> tr/a-z/A-Z/   |> p  # HELLO WORLD
  
  # m//g extracts all matches as an array:
  "foo123bar456" |> /\d+/g |> p  # 123 456
  
  # full pipeline: read input, strip newlines, split, count word frequencies
  # man ls | stryke 'input |> s@\n@@g |> split |> frequencies |> ddump |> p'
  
  # extract all emails from text, deduplicate
  # cat log.txt | stryke 'input |> /[\w.]+@[\w.]+/g |> distinct |> e p'
  
  # capture groups with /g:
  "a=1 b=2" |> /(\w+)=(\w+)/g |> ddump |> p
  

  Pipeline builtins — designed for |> chains:

  # ── input / output ─────────────────────────────────────────────────
  input                                # slurp all of stdin as one string
  input($fh)                           # slurp a filehandle
  # cat data.txt | stryke 'input |> lines |> e p'
  
  # ── string → list ──────────────────────────────────────────────────
  "hello\nworld" |> lines |> ddump |> p  # ("hello", "world")
  "foo bar baz"  |> words |> ddump |> p  # ("foo", "bar", "baz")
  "hello"        |> chars |> ddump |> p  # ("h","e","l","l","o")
  "  hello  "    |> trim  |> p  # "hello"
  
  # ── case conversion ────────────────────────────────────────────────
  "helloWorld"     |> snake_case  |> p  # hello_world
  "hello_world"    |> camel_case  |> p  # helloWorld
  "Hello World"    |> kebab_case  |> p  # hello-world
  
  # ── aggregation / stats ────────────────────────────────────────────
  1 .. 100 |> avg    |> p  # 50.5
  1 .. 100 |> stddev |> p  # 28.86607…
  "hello"  |> chars  |> frequencies |> ddump |> p
  # { h => 1, e => 1, l => 2, o => 1 }
  
  # ── frequencies + top ──────────────────────────────────────────────
  "the quick brown fox" |> chars |> frequencies |> top 3 |> ddump |> p
  # top 3 chars by count
  
  # ── count_by { BLOCK } LIST ────────────────────────────────────────
  1 .. 20 |> count_by { $_ % 2 == 0 ? "even" : "odd" } |> ddump |> p
  # { odd => 10, even => 10 }
  
  # ── numeric transforms ─────────────────────────────────────────────
  1 .. 10  |> clamp 3, 7    |> ddump |> p  # 3 3 3 4 5 6 7 7 7 7
  1 .. 5   |> normalize     |> ddump |> p  # 0 0.25 0.5 0.75 1
  
  # ── inverse grep (regex) ───────────────────────────────────────────
  1 .. 10 |> grep_v "^[35]$" |> ddump |> p  # removes 3 and 5
  
  # ── hash manipulation ──────────────────────────────────────────────
  val $h = {a => 1, b => 2, c => 3}
  $h |> select_keys "a", "c" |> ddump |> p  # { a => 1, c => 3 }
  
  # ── pluck key from list of hashrefs ────────────────────────────────
  val @people = ({name=>"Alice",age=>30}, {name=>"Bob",age=>25})
  @people |> pluck "name" |> ddump |> p  # ("Alice", "Bob")
  
  # ── serialization ──────────────────────────────────────────────────
  val $data = {a => 1, b => [2,3]}
  $data |> to_json |> p  # {"a":1,"b":[2,3]}
  @people |> to_csv |> p  # CSV with headers
  val $cfg = {title => "My App", package => {name => "myapp", version => "1.0"}}
  $cfg |> to_toml |> p  # TOML with [package] table
  $data |> to_yaml |> p  # YAML with --- header
  $data |> to_xml  |> p  # XML with <root> wrapper
  fr |> map +{name => $_, size => format_bytes(size)} |> th |> to_file("report.html") |> xopen  # cyberpunk HTML table → browser
  fr |> map +{name => $_, size => format_bytes(size)} |> tmd |> to_file("report.md") |> xopen  # GFM Markdown table → viewer
  # same pipelines in ~> syntax:
  ~> fr map +{name => $_, size => format_bytes(size)} th to_file($_, "report.html") xopen
  ~> fr map +{name => $_, size => format_bytes(size)} tmd to_file($_, "report.md") xopen
  fr |> map +{name => $_, size => format_bytes(size)} |> tbl |> p                      # plain-text aligned table
  fr |> map +{name => $_, size => format_bytes(size)} |> tmd |> clip                   # markdown table → clipboard
  
  # ── data visualization ─────────────────────────────────────────────
  # sparkline — inline Unicode trend line from numbers
  (3,7,1,9,4,2,8,5) |> spark |> p  # ▃▆▁█▄▂▇▅
  map { int(rand(100)) } 1..20 |> spark |> p  # random sparkline
  
  # bar_chart (bars) — horizontal colored bars from hashref
  qw(a b a c a b) |> freq |> bars |> p  # word frequency bars
  cat("Cargo.toml") |> words |> freq |> bars |> p  # word freq from file
  fr |> map { path_ext($_) } |> freq |> bars |> p  # file extension breakdown
  
  # histo — vertical histogram, top N by count
  cat("Cargo.toml") |> chars |> freq |> histo |> p  # character distribution
  map { int(rand(10)) } 1..100 |> freq |> histo |> p  # dice roll distribution
  
  # to_table (tbl) — plain-text column-aligned table with box drawing
  qw(a b a c a b) |> freq |> tbl |> p  # freq as table
  fr |> map +{name => $_, size => format_bytes(size)} |> tbl |> p  # file listing table
  fr |> map +{name => $_, ext => path_ext($_)} |> tbl |> p  # files with extensions
  
  # flame — terminal flamechart from nested hashrefs
  flame({main => {parse => 30, eval => {compile => 15, run => 45}}, init => 10}) |> p
  cat("Cargo.toml") |> chars |> freq |> flame |> p  # flat flame from char freq
  
  # gauge — single-value progress bar with color coding
  p gauge(0.73)  # [██████████████████████░░░░░░░░] 73%
  p gauge(45, 100)  # value/max form
  fr |> cnt |> gauge($_, 500) |> p  # file count vs target
  
  # spinner — animated braille spinner while block runs
  val $r = spinner "loading" { sleep 2; 42 }  # returns block result
  val $data = spinner "fetching" { fetch_json($url) }  # wrap any slow operation
  # spinner_start / spinner_stop — manual control for multi-step work
  val $s = spinner_start("processing")
  do_step1(); do_step2(); do_step3()
  spinner_stop($s)
  
  # clip — copy pipeline output to clipboard
  fr |> map +{name => $_, size => format_bytes(size)} |> tmd |> clip  # markdown table → clipboard
  cat("Cargo.toml") |> words |> freq |> tbl |> clip  # table → clipboard
  
  # combine charts: same data, multiple views
  val %f = %{cat("Cargo.toml") |> words |> freq}
  %f |> bars |> p  # horizontal bars
  %f |> histo |> p  # vertical histogram
  %f |> tbl |> p  # aligned table
  %f |> flame |> p  # flamechart
  values %f |> spark |> p  # inline sparkline
  
  # ~> syntax equivalents — no |> needed
  ~> qw(a b a c a b) freq bars p
  ~> qw(a b a c a b) freq histo p
  ~> qw(a b a c a b) freq tbl p
  ~> (3,7,1,9,4) spark p
  
  # ── inline ANSI colors ─────────────────────────────────────────────
  p "#{red}ERROR#{off} #{green_bold}OK#{off}"  # color names as #{} builtins
  p "#{rgb(255,100,0)}ORANGE#{off}"  # true color (24-bit)
  p "#{color256(196)}RED#{off}"  # 256-color palette
  
  # ── stringify / str — parseable stryke literals ──────────────────────
  $data |> str |> p  # +{a => 1, b => [2, 3]}
  val $fn = fn { $_ * 2 }
  $fn |> str |> p  # fn { $_ * 2; }
  range(1, 3) |> str |> p  # (1, 2, 3)
  # round-trip: str -> eval -> callable
  val $f = fn ($x: Int) { $x + 1 }
  val $f2 = $f |> str |> eval
  $f2->(5) |> p  # 6
  
  # ── partition / min_by / max_by / zip_with ─────────────────────────
  val ($yes, $no) = partition { $_ > 5 } 1..10
  val $smallest = min_by { length } @words
  val $largest  = max_by { length } @words
  val @sums = zip_with { $_0 + $_1 } [1,2,3], [10,20,30]  # 11 22 33
  
  # ── pretty-print (indented dump) ───────────────────────────────────
  val $nested = {key => [1, {nested => "val"}]}
  $nested |> ddump |> p
  
  # ── write to file (returns content for further piping) ─────────────
  val $text = "hello\nworld\n"
  $text |> to_file "/tmp/out.txt"
  
  # ── file I/O helpers ────────────────────────────────────────────────
  val @lines = read_lines "/tmp/out.txt"  # slurp file → list of lines
  append_file "/tmp/out.txt", "extra\n"  # append to file
  val $tmp = tempfile()  # create temp file, returns path
  val $dir = tempdir()  # create temp directory, returns path
  
  # ── JSON file I/O ──────────────────────────────────────────────────
  write_json "/tmp/data.json", {a => 1, b => 2}  # write hash as JSON file
  val $obj = read_json "/tmp/data.json"  # read JSON file → hashref
  
  # ── interleave ─────────────────────────────────────────────────────
  val @merged = interleave [1,2,3], [10,20,30]  # (1,10,2,20,3,30)
  
  # ── glob_match / which_all ──────────────────────────────────────────
  p glob_match "*.txt", "readme.txt"  # 1 (matches)
  val @bins = which_all "perl"  # all paths for "perl" in $PATH
  
  # ── zsh glob qualifiers — world's first in a scripting language ────
  # Stryke imports the full zshrs glob engine (zsh-compatible). Every
  # builtin that accepts a glob — `glob`, `glob_par`, `slurp`/`c`/`cat`,
  # `swallow`/`swa`, `ingest`/`ing`, `pwatch`, `par_find_files`,
  # `<*.txt>`, … — applies the qualifiers without a single line of
  # stryke-side parsing. Source of truth is `zsh::glob` from `../zshrs`.
  val @dirs = glob "**(/)"          # directories only, recursive
  val @files = glob "**(.)"         # regular files only, recursive
  val @links = glob "**(@)"         # symlinks only
  val @exec = glob "**(*)"          # executable files
  val @big = glob "**(L+1024)"      # files larger than 1024 bytes
  val @recent = glob "**(om[1])"    # most recently modified, take 1
  val @safe = glob "doesnotexist*(N)"  # NULL_GLOB — empty list, no error
  
  # `c()` is a slurp. A plain wildcard sweep skips matches it can't read as
  # a regular file (dangling symlink, vanished file, directory) the way
  # `grep -r` does — one bad entry never aborts the sweep:
  c "**.rs"   # OK: every .rs file's contents; incidental bad matches skipped
  # A qualifier glob is an explicit selection, so it stays strict — asking to
  # read a directory is always a bug:
  c "**(.)"   # OK: concatenated contents of every file recursively
  c "**(/)"   # ERROR: "slurp: not a regular file: ./sub"
  
  # `swallow` is the per-file hash sibling of `slurp` — returns
  # `{ canonical_abspath => raw_bytes }`. Always bytes (works for binary
  # files); keys are absolute paths with symlinks flattened via
  # `fs::canonicalize`. Same match policy as slurp: a plain sweep skips
  # bad references (dangling symlinks, vanished files, dirs); a qualifier
  # glob stays strict.
  val %src = swallow "src/**/*.rs"   # every Rust source file, raw bytes
  val %imgs = swa "assets/**/*.{png,jpg}"
  val %safe = swallow "missing*(N)"  # (N) null-glob → empty hash
  
  # `ingest` is the streaming variant of `swallow` — yields
  # `[abspath, bytes]` one file at a time so only one file's bytes are
  # resident at any moment. Same eager glob expansion + match policy
  # (plain sweep skips bad references, qualifier glob strict), but file reads are
  # deferred to each iteration step. For-loops over an ingest iterator
  # pull lazily (no `to_list()` materialisation); use `|>` pipes or
  # explicit `->next` driving the same way.
  for val $pair (ingest "data/**/*.bin") {
      val ($path, $bytes) = @$pair
      # process one file's bytes, then they go out of scope
  }
  val $it = ing "logs/*.log"; while (val $p = $it->next->[0]) { ... }
  
  # `burp` is the inverse of `swallow` — take a `{path => content}` hash
  # and write each entry. Parent directories are created automatically,
  # so the canonical swallow → mutate → burp round-trip works even when
  # the destination tree doesn't yet exist. Pass via hashref (`\%h` or
  # inline `{ ... }`); returns the integer count of files written.
  var %src = swallow "src/**/*.rs"
  for val $p (keys %src) { $src{$p} = uc $src{$p} }
  val $n = burp \%src                          # in-place update
  burp { "out/README.md" => "# Hello\n", "out/src/main.rs" => "..." }
  

  Full qualifier reference — stryke supports every zsh glob qualifier (man zshexpn, Filename Generation > Glob Qualifiers), inherited verbatim from zsh::glob:

  Family	Qualifier	Match
  type	(/) (.) (@) (=) (p) (%b) (%c) (%) (*)	dir / regular file / symlink / socket / FIFO / block-dev / char-dev / any-dev / executable
  perm (EUID)	(r) (w) (x)	readable / writable / executable
  perm (other)	(R) (W) (X) (A) (I) (E)	world r/w/x · group r/w/x
  special	(s) (S) (t)	setuid / setgid / sticky
  mode bits	(f<bits>)	exact mode-bit match, e.g. (f644)
  owner	(U) (G) (u<N>) (g<N>)	EUID / EGID / numeric uid / numeric gid
  device	(d<N>)	match by device number
  size	(L[unit]±N)	bytes default; units p k m g t; +N greater, -N less
  links	(l±N)	hard-link count
  times	(a±N) (m±N) (c±N)	atime / mtime / ctime; units s m h d w M
  sort	(o…) (O…)	asc / desc on n L l a m c d; (oN) no-sort
  select	([N]) ([N,M])	Nth / slice; combine with sort, e.g. (om[1]) newest
  flags	(N) (D) (F) (M) (T) (n)	null-glob / include dotfiles / non-empty dir / mark-dirs / list-types / numeric-sort
  eval	(e'CMD') (+func)	shell-eval predicate / function-as-test
  join	(P…) (Q…)	prefix / postfix join words around each match
  colon	(:s/PAT/REPL/) (:e) (:r) (:t) (:h) …	sed-style + tail/root/extension/head modifiers on each result path
  combinators	^ - , chain	negate / toggle follow-symlinks / OR / chained-AND

  Blockless |> rules for grep/filter: string literals test $_ eq EXPR, numbers test $_ == EXPR, regexes test $_ =~ EXPR, anything else (e.g. defined) uses standard Perl grep semantics (sets $_, evaluates expression).

  Coderef-in-block-position — wherever a { BLOCK } is accepted (grep, map, sort, first, any, all, none, take_while, drop_while, reject, partition, min_by, max_by, plus their pipe-forward variants), a coderef-shaped expression also works directly. Runtime check: if the EXPR evaluates to a code ref, it is called with the current element(s) as positional args; otherwise the value's truthiness drives filtering (or its result becomes the mapped value, comparator integer, etc.). Eliminates the { $f($_) } / { $f->($_) } boilerplate.

  val $is_big = fn ($x) { $x > 3 }
  val @r = grep $is_big, @l                # was: grep { $is_big->($_) } @l
  val @r = @l |> grep $is_big              # pipe-forward variant
  val @r = first $is_big, @l               # tier-2 builtin, no parens, no block
  val @r = take_while $is_big, @l
  
  # Sort comparators receive ($a, $b) positionally — no $a/$b global magic:
  val $cmp = fn ($a, $b) { $b <=> $a }     # or fn { _0 <=> _1 } using positional aliases
  val @s = sort $cmp @l                    # descending
  

  Threading (~>) excluded — whitespace-delimited stages can't disambiguate ~> @l grep $f from "two stages", so threading still requires { $f(_) }. Use |> for the bare-coderef form, or stay with { } blocks under ~>.

  Under --compat: dispatch is skipped, restoring Perl's "evaluate EXPR per element, filter by truthiness" semantics. A coderef value is always truthy, so grep $f, @l keeps every element under --compat.

  Precedence: |> binds looser than || but tighter than ?: / and/or/not — the slot sits between parse_ternary and parse_or_word in the parser stack. So $x + 1 |> f parses as f($x + 1), and 0 || 1 |> yes parses as yes(0 || 1). The RHS must be a call, builtin, method invocation, bareword, or coderef expression; bare binary expressions / literals on the right are a parse error (42 |> 1 + 2 is rejected).

• ~> macro (thread, t, ->>) — Clojure-inspired threading macro for clean multi-stage pipelines without repeating |>. Stages are bare function names, functions with blocks, parenthesized calls name(args) where $_ (or bare _) is the threaded-value placeholder (must appear at least once in args, can sit in any position — first, last, middle, nested), or anonymous blocks (>{} / fn {}). Use |> after ~> to continue piping. Blocks can use bare _ for maximum conciseness — map{_*2} is equivalent to map{$_ * 2}.

  # ultra-concise — bare _ eliminates sigil noise
  ~>1:10map{_*2}fi{_>5}sum p                          # 104
  
  # ~> shines with multiple block-taking functions — no |> repetition
  @data = 1..20
  ~> @data grep{_ % 2 == 0} map{_ * _} sort{$_1 <=> $_0} |> join "," |> p
  # 400,324,256,196,144,100,64,36,16,4
  
  # Compare: same pipeline with |> requires more syntax
  @data |> grep{_ % 2 == 0} |> map{_ * _} |> sort{$_1 <=> $_0} |> join "," |> p
  
  # Long data processing pipeline
  @nums = 1..100
  ~> @nums grep{_ % 3 == 0} map{_ * 2} grep{_ > 50} sort{$_1 <=> $_0} |> head 5 |> join "," |> p
  # 198,192,186,180,174
  
  # Anonymous blocks for custom transforms
  ~> 100 >{_ / 2} >{_ + 10} >{_ * 3} p  # 180
  
  # Process list of hashes
  @users = ({name=>"alice",age=>30}, {name=>"bob",age=>25}, {name=>"carol",age=>35})
  ~> @users sort{$_0->{age} <=> $_1->{age}} map{_->{name}} |> join "," |> p
  # bob,alice,carol
  
  # String processing with unary builtins
  ~> "  hello world  " trim uc p                 # HELLO WORLD
  
  # Parenthesized call stages — `_` or `$_` is the threaded-value placeholder
  fn add2 { $_0 + $_1 }
  ~> 10 add2(_, 5) p                              # add2(10, 5)        => 15
  ~> 10 add2(5, _) p                              # add2(5, 10)        => 15  (any position)
  ~> 10 add2(_, 5) add2(_, 100) p                 # chains: 15 then 115
  fn add3 { $_0 + $_1 + $_2 }
  ~> 10 add3(5, _, 10) p                          # add3(5, 10, 10)    => 25
  # `_` works inside nested expressions too:
  fn mul { $_0 * $_1 }
  ~> 10 mul(_ + 1, 2) p                           # mul(11, 2)         => 22
  
  # Reduce with $_0/$_1
  ~> (1..10) reduce { $_0 + $_1 } p              # 55
  
  # Sort and unique
  @data = (3,1,4,1,5,9,2,6,5,3)
  ~> @data sort { $_0 <=> $_1 } uniq |> join "," |> p   # 1,2,3,4,5,6,9
  

  When to use ~> vs |>:

  • ~>: Best for chains of block-taking functions (map { }, grep { }, sort { }, reduce { })
  • |>: Best for blockless expressions (map $_ * 2, grep $_ > 5) and unary functions

  # |> with blockless expressions — cleanest for simple transforms
  1..20 |> grep $_ % 2 == 0 |> map $_ * $_ |> grep $_ > 50 |> join "," |> p
  # 64,100,144,196,256,324,400
  
  # ~> with blocks — cleanest when every stage needs a block
  ~> @data map { complex($_) } grep { validate($_) } sort { $_0 cmp $_1 } |> p
  

  Stage types:

  • Bare function: ~> "hello" uc trim — applies unary builtins in sequence
  • Function with block: ~> @data map{_ * 2} grep{_ > 5} — block-taking functions (bare _ or $_)
  • Anonymous block: ~> 5 >{_ * 2} or fn { } — custom transforms

  Termination: |> ends the ~> macro: ~> @l f1 f2 f3 |> f4 parses as (~> @l f1 f2 f3) |> f4.

  Numeric/statistical pipelines:

  # Sum of squares of even numbers 1-10
  ~> (1..10) grep{_ % 2 == 0} map{_ * _} sum p                # 220
  
  # Mean of squares
  ~> (1..10) map{_ * _} mean p                                 # 38.5
  
  # Multiples of 7 up to 100, doubled, summed
  ~> (1..100) grep{_ % 7 == 0} map{_ * 2} sum p               # 1470
  
  # Sum of odd squares, sqrt, truncate
  ~> (1..50) grep{_ % 2 == 1} map{_ ** 2} sum sqrt int p      # 144
  
  # Factorial via product
  ~> (1..10) product p                                        # 3628800
  
  # Remove duplicates, then sum
  ~> (1,1,2,2,3,3,4,5,5) uniq sum p                           # 15
  
  # Shuffle, dedupe, sum (same result, random order internally)
  ~> (1..20) shuffle uniq sum p                               # 210
  
  # Statistical measures
  ~> (1..10) mean p                                           # 5.5
  ~> (1..10) median p                                         # 5.5
  ~> (1..10) stddev p                                         # 2.87228...
  

  String pipelines:

  # Full transformation
  ~> " hello world " trim uc rev lc ucfirst snake_case camel_case kebab_case to_json p
  # "d-lrow-olleh"
  
  # String list operations
  ~> ("apple","banana","cherry","date") shuffle rev minstr p  # apple
  ~> ("apple","banana","cherry","date") shuffle rev maxstr p  # date
  

  Sorting and aggregation:

  # Sort then get min/max
  ~> (5,2,8,1,9,3) sort { $_0 <=> $_1 } min p                 # 1
  ~> (5,2,8,1,9,3) sort { $_0 <=> $_1 } max p                 # 9
  
  # Pairs: extract keys and values
  ~> (1,2,3,4,5,6) pairkeys |> join "," |> p                  # 1,3,5
  ~> (1,2,3,4,5,6) pairvalues |> join "," |> p                # 2,4,6
  

  Compare with |> syntax (same result, more typing):

  # ~> version (bare _)
  ~> (1..10) grep{_ % 2 == 0} map{_ * _} sum p
  
  # |> version
  (1..10) |> grep{_ % 2 == 0} |> map{_ * _} |> sum |> p
  

  Language comparison — the same 10-stage pipeline:

  # stryke: 1 line, reads left-to-right, no noise
  ~> " hello world " trim uc rev lc ucfirst snake_case camel_case kebab_case to_json p
  

  # Perl 5: needs CPAN modules, verbose method chains
  use String::CamelCase qw(camelize decamelize)
  use JSON
  var $s = " hello world "
  $s =~ s/^\s+|\s+$//g  # trim
  $s = uc($s)
  $s = reverse($s)
  $s = lc($s)
  $s = ucfirst($s)
  $s =~ s/([A-Z])/_\l\$1/g; $s =~ s/^_//  # snake_case (manual)
  $s = camelize($s)  # camel_case (CPAN)
  $s =~ s/([A-Z])/-\l\$1/g; $s =~ s/^-//  # kebab_case (manual)
  print encode_json($s), "\n"
  

  // JavaScript: no built-in case converters, needs helper functions
  const snakeCase = s => s.replace(/([A-Z])/g, '_\$1').toLowerCase().replace(/^_/, '');
  const camelCase = s => s.replace(/_([a-z])/g, (_, c) => c.toUpperCase());
  const kebabCase = s => s.replace(/([A-Z])/g, '-\$1').toLowerCase().replace(/^-/, '');
  const ucfirst = s => s.charAt(0).toUpperCase() + s.slice(1);
  const rev = s => s.split('').reverse().join('');
  
  let s = " hello world ";
  s = s.trim();
  s = s.toUpperCase();
  s = rev(s);
  s = s.toLowerCase();
  s = ucfirst(s);
  s = snakeCase(s);
  s = camelCase(s);
  s = kebabCase(s);
  console.log(JSON.stringify(s));
  

  # Python 3: no built-in case converters, needs helper functions
  import json
  import re
  
  def snake_case(s): return re.sub(r'([A-Z])', r'_\1', s).lower().lstrip('_')
  def camel_case(s): return re.sub(r'_([a-z])', lambda m: m.group(1).upper(), s)
  def kebab_case(s): return re.sub(r'([A-Z])', r'-\1', s).lower().lstrip('-')
  
  s = " hello world "
  s = s.strip()
  s = s.upper()
  s = s[::-1]
  s = s.lower()
  s = s[0].upper() + s[1:]  # ucfirst
  s = snake_case(s)
  s = camel_case(s)
  s = kebab_case(s)
  print(json.dumps(s))
  

  stryke: 1 line. Perl 5: 10+ lines + CPAN. JavaScript: 15+ lines. Python: 15+ lines.

  Lisp hell — without |>, the same pipeline becomes unreadable:

  # stryke with |> : reads left-to-right
  " hello world " |> trim |> uc |> rev |> lc |> ucfirst |> rev |> snake_case |> camel_case |> kebab_case |> rev |> uc |> lc |> trim |> to_json |> p
  # "d-lrow-olleh"
  
  # Without |> : nested calls, reads inside-out (lisp hell)
  p(to_json(trim(lc(uc(rev(kebab_case(camel_case(snake_case(rev(ucfirst(lc(rev(uc(trim(" hello world ")))))))))))))))
  

  The pipe-forward operator eliminates the cognitive overhead of matching parentheses and reading inside-out.

• Short aliases — 1-3 character aliases for common functions, designed for ~>/|> pipelines:

  # Long form
  ~> " hello world " trim uc rev lc ucfirst snake_case camel_case kebab_case to_json p
  
  # Short form (same result)
  ~> " hello world " tm uc rv lc ufc sc cc kc tj p
  

  Alias	Function	Alias	Function	Alias	Function
  Thread/Pipe		String		Case
  ~>	thread	tm	trim	sc	snake_case
  p	len	length	cc	camel_case
  pr	print	ufc	ucfirst	kc	kebab_case
  		lfc	lcfirst	qm	quotemeta
  List		rev
  gr	grep	ch	chars	Serialize
  so	sort	ln	lines	tj	to_json
  rd	reduce	wd	words	ty	to_yaml
  hd	head/take			tt	to_toml
  tl	tail	Unique/Dedup		tc	to_csv
  drp	drop/skip	uq	uniq	tx	to_xml
  fl	flatten	dup	dedup	th	to_html
  cpt	compact	shuf	shuffle	tmd	to_markdown
  				dd	ddump
  				xo	xopen
  cat	slurp			Deserialize
  il	interleave	Stats		jd	json_decode
  en	enumerate	sq	sqrt	yd	yaml_decode
  wi	with_index	med	median	td	toml_decode
  chk	chunk	std	stddev	xd	xml_decode
  zp	zip	var	variance	je	json_encode
  fst	first	clp	clamp	ye	yaml_encode
  frq	frequencies	nrm	normalize	te	toml_encode
  win	windowed			xe	xml_encode
  		Crypto
  File/Path		s1	sha1	Encoding
  sl	slurp	s256	sha256	b64e	base64_encode
  wf	write_file	m5	md5	b64d	base64_decode
  rl	read_lines	uid	uuid	hxe	hex_encode
  rb	read_bytes			hxd	hex_decode
  swa	swallow
  ing	ingest
  	burp
  af	append_file	HTTP		ue	url_encode
  rj	read_json	ft	fetch	ud	url_decode
  wj	write_json	ftj	fetch_json	gz	gzip
  bn	basename	fta	fetch_async	ugz	gunzip
  dn	dirname	hr	http_request	zst	zstd
  rp	realpath	pft	par_fetch	uzst	zstd_decode
  wh	which
  pwd	getcwd	CSV/Data		DateTime
  tf	tempfile	cr	csv_read	utc	datetime_utc
  tdr	tempdir	cw	csv_write	now	datetime_now_tz
  hn	gethostname	pcr	par_csv_read	dte	datetime_from_epoch
  el	elapsed	df	dataframe	dtf	datetime_strftime
  def	defined	sql	sqlite
  rss	proc_mem

• fn keyword — alias for sub. Both fn name { } and fn { } work identically to sub.

  fn double($x) { $x * 2 }
  p double(21)                    # 42
  
  val $f = fn { _ * 2 }
  p $f->(21)                      # 42
  
  # implicit zero-arg coderef — at top-level, RHS starting with bare `_` / `_N`
  # auto-wraps as `fn { ... }`. Inside any block, `_` is still the topic.
  val $g = _ * 2
  p $g->(21)                      # 42
  val $h = _ + _1
  p $h->(3, 4)                    # 7
  

• Closure arguments $_0, $_1, ... $_N — numeric closure arguments inspired by Swift. All arguments passed to any fn (named or anonymous) are available as $_0 (first), $_1 (second), $_2 (third), up to $_N for any number of arguments. These work alongside or instead of Perl's @_, $_, $a, $b. Both $_, bare _, and $_0 refer to the first argument — _ * 2, $_ * 2, and $_0 * 2 are all equivalent. Use bare _ for maximum conciseness in blocks.

  # $_0 in |> pipes (single-arg: $_0 == $_)
  (1..5) |> map { $_0 * 2 } |> join "," |> p           # 2,4,6,8,10
  (1..10) |> grep { $_0 % 2 == 0 } |> sum |> p         # 30
  
  # $_0/$_1 in |> pipes (two-arg: $_0/$_1 == $a/$b)
  (5,2,8,1) |> sort { $_0 <=> $_1 } |> join "," |> p   # 1,2,5,8
  (1..5) |> reduce { $_0 + $_1 } |> p                  # 15
  (1..5) |> reduce { $_0 * $_1 } |> p                  # 120 (factorial)
  ("banana","apple","cherry") |> sort { length($_0) <=> length($_1) } |> join "," |> p  # apple,banana,cherry
  
  # $_0/$_1 in ~> macro
  ~> (1..5) map { $_0 * 2 } sum p                  # 30
  ~> (1..5) reduce { $_0 + $_1 } p                 # 15
  ~> (1..5) reduce { $_0 * $_1 } p                 # 120
  ~> (5,2,8,1) sort { $_0 <=> $_1 } |> join "," |> p  # 1,2,5,8
  ~> (1..10) grep { $_0 % 2 == 0 } map { $_0 * $_0 } sum p  # 220
  
  # Multi-arg anonymous subs: $_0, $_1, ... $_N
  val $add3 = fn { $_0 + $_1 + $_2 }
  p $add3->(1, 2, 3)  # 6
  
  val $mul5 = fn { $_0 * $_1 * $_2 * $_3 * $_4 }
  p $mul5->(1, 2, 3, 4, 5)  # 120
  
  val $concat = fn { "$_0-$_1-$_2-$_3" }
  p $concat->("a", "b", "c", "d")  # a-b-c-d
  
  # Direct access via @_ still works
  val $join_args = fn { join("-", @_) }
  p $join_args->("x", "y", "z")  # x-y-z
  
  # Using $_0 closures with |> pipes
  val $double = fn { $_0 * 2 }
  val $triple = fn { $_0 * 3 }
  5 |> $double |> $triple |> p               # 30
  
  # Using $_0/$_1 closures in reduce
  val $add = fn { $_0 + $_1 }
  (1..5) |> reduce { $add->($_0, $_1) } |> p # 15
  
  # Using $_0/$_1/$_2 closure
  val $mul3 = fn { $_0 * $_1 * $_2 }
  p $mul3->(2, 3, 4)  # 24
  
  # Using $_0/$_1 closure as comparator
  val $cmp = fn { $_0 <=> $_1 }
  (5,2,8,1) |> sort { $cmp->($_0, $_1) } |> join "," |> p  # 1,2,5,8
  
  # User-defined functions in ~> (bare stage, no block needed)
  fn double { $_0 * 2 }
  fn triple { $_0 * 3 }
  fn add5   { $_0 + 5 }
  fn square { $_0 ** 2 }
  fn half   { $_0 / 2 }
  ~> 2 double triple add5 square half p  # 144.5
  
  fn inc  { $_0 + 1 }
  fn dec  { $_0 - 1 }
  fn dbl  { $_0 * 2 }
  fn neg  { -$_0 }
  fn abs_ { abs($_0) }
  ~> 5 inc dbl dec neg abs_ dbl inc p    # 23
  
  fn wrap  { "[$_0]" }
  fn upper { uc($_0) }
  fn trim_ { trim($_0) }
  fn rev_  { rev($_0) }
  fn bang  { "$_0!" }
  ~> "  hello  " trim_ upper rev_ wrap bang p  # [OLLEH]!
  
  # User-defined functions inside blocks
  fn is_even { $_0 % 2 == 0 }
  ~> (1..10) grep{is_even(_)} sum p  # 30
  
  ~> (1..5) map{square(_)} sum p     # 55
  
  # Multi-arg user-defined functions
  fn add  { $_0 + $_1 }
  fn mul3 { $_0 * $_1 * $_2 }
  p add(3, 4)                                # 7
  p mul3(2, 3, 4)                            # 24
  
  # Inline transforms with >{ } (arrow block)
  ~> 5 >{_ * 2} >{_ + 10} p               # 20
  ~> 100 >{_ / 2} >{_ + 10} >{_ * 3} p    # 180
  

• Block params { |$var| body } — name the block's implicit arguments with Ruby-style |$params| at the start of a block. For single-param blocks (map, grep, each), the param aliases $_. For two-param blocks (sort, reduce), they alias $a/$b. For N≥3 params, they alias $_, $_1, $_2, etc.

  # Single param — aliases $_
  map { |$n| $n * $n }, 1..5                         # 1 4 9 16 25
  grep { |$x| $x > 3 }, 1..6                         # 4 5 6
  (1..3) |> map { |$n| $n + 10 } |> join ","         # 11,12,13
  
  # Two params — aliases $a/$b
  sort { |$x, $y| $y <=> $x }, 3, 1, 4, 1, 5        # 5 4 3 1 1
  reduce { |$acc, $val| $acc + $val }, 1..10         # 55
  

stryke is not a full perl replacement: many real .pm files (especially XS modules) will not run. See PARITY_ROADMAP.md.

---

[0x08a] --no-interop MODE

--no-interop is the idiomatic-stryke-only mode: every Perl-ism that has a stryke replacement becomes a parse-time error so codebases stay on the stryke side of the language. Cargo-cult Perl idioms can't sneak in, and grep'ing for \bscalar\b / \bsub\b / \bsay\b in your sources stays signal-only.

Perl-ism	Rejected with	Use instead
sub NAME { … } / sub { … }	stryke uses 'fn' instead of 'sub' (--no-interop)	fn NAME { … } / fn { … }
say EXPR	stryke uses 'p' instead of 'say' (--no-interop)	p EXPR
reverse EXPR	stryke uses 'rev' instead of 'reverse' (--no-interop)	rev EXPR (works for both strings and lists)
$a / $b outside sort/reduce blocks	stryke uses '$_0' / '$_1' instead of '$a' (--no-interop)	$_0 / $_1 positional block params
scalar EXPR (any form)	stryke uses 'len' (also 'cnt' / 'count') instead of 'scalar' (--no-interop)	see the scalar mapping below

scalar mapping under --no-interop

scalar was overloaded with at least four distinct semantics in Perl. Under --no-interop each idiom has its own verb so the meaning is explicit at the call site:

Perl idiom	What it does	Stryke spelling
scalar @a / scalar(@$ref) / scalar @{$r}	element count	len @a / len(@$ref) / len @{$r} (aliases cnt, count) — compiles to Op::ArrayLen / Op::ArrayDerefLen fast path
scalar keys %h / scalar values %h	key / value count	len keys %h / len values %h
scalar grep { … } @a / scalar split(…) / scalar qw(…)	match / part / element count	len grep { … } @a / len split(…) / len qw(…)
scalar reverse $s	string reverse (vs. list reverse)	rev $s
scalar unpack(FMT, STR)	first decoded element	unpack_first(FMT, STR) (aliases unpack1, up1) — equivalent to head(unpack(…))
scalar splice(@a, off, n)	last removed element	splice_last(@a, off, n) (aliases splice1, spl_last) — desugars to tail(splice(@a, off, n)), mutates in place
scalar \cmd``	joined stdout as string	already the default — stryke backticks return a single string regardless of context, so no spelling change needed
scalar %h (Perl's hash-fill diagnostic, e.g. "3/8")	dead semantics — never load-bearing	use len keys %h for the count

Default mode (no --no-interop) still accepts every Perl-ism listed above for compat with stock .pm / .pl sources.

---

[0x08b] STRING COORDINATES — BYTES VS CODEPOINTS

Stryke runs string code in two coordinate systems. Perl 5 builtins stay byte-indexed for binary-protocol / .pm-source compat. Stryke extensions are codepoint-indexed so search positions and slice bounds line up. They are never auto-converted — pick one coordinate system per expression, keep it consistent, and never mix a byte-position output into a codepoint-position input.

Operation	Stryke (codepoints)	Perl 5 (bytes)
Length	len $s	length $s
Index	$s[$i]	substr $s, $i, 1
Slice	$s[$a:$b] (inclusive both ends)	substr $s, $start, $len
Search forward	cindex $s, $needle [, $from]	index $s, $needle [, $from]
Search backward	crindex $s, $needle [, $from]	rindex $s, $needle [, $from]
Match position	—	pos $s (regex \G anchor)

Concrete example:

val $s = "hello ─ world"     # `─` is 3 bytes / 1 codepoint
length $s                    # 15 (bytes)
len    $s                    # 13 (codepoints)
index  $s, "world"           # 9  (byte position — past the 3-byte `─`)
cindex $s, "world"           # 7  (codepoint position)
$s[7]                        # "w"
substr $s, 9, 1              # "w"
$s[7:11]                     # "world"  — codepoint slice
substr $s, 9, 5              # "world"  — byte substr

Rule: never feed an index / pos / length result into a [$a:$b] slice, and never feed a cindex / crindex / len result into substr / index. The coordinate systems silently misalign on any string containing non-ASCII bytes.

--no-interop does not force this split — both systems remain available because Perl 5 binary-protocol code legitimately needs byte positions. The split is a coordinate-system choice, not a stylistic one.

---

[0x09] ARCHITECTURE

 ┌─────────────────────────────────────────────────────┐
 │  Source ──▶ Lexer ──▶ Parser ──▶ AST                │
 │                                    │                │
 │                                    ▼                │
 │                            Compiler (compiler.rs)   │
 │                                    │                │
 │                                    ▼                │
 │                            Bytecode (bytecode.rs)   │
 │                                    │                │
 │                    ┌───────────────┴───────────┐    │
 │                    ▼                           ▼    │
 │               VM (vm.rs)                 Cranelift  │
 │                    │                      Block JIT │
 │                    ▼                                │
 │       Rayon work-stealing scheduler                 │
 │       CORE 0 │ CORE 1 │ ... │ CORE N                │
 └─────────────────────────────────────────────────────┘

• Lexer (strykelang/lexer.rs) — context-sensitive tokenizer for Perl's ambiguous syntax (regex vs division, hash vs modulo, heredocs, interpolation).
• Parser (strykelang/parser.rs) — recursive descent + Pratt precedence climbing.
• Compiler / VM (strykelang/compiler.rs, strykelang/vm.rs) — 100% lowered to bytecode. Compiled subs use slot ops for frame-local my scalars (O(1)). Lowering covers BEGIN/UNITCHECK/CHECK/INIT/END with Op::SetGlobalPhase, varsync, tie, scalar compound assigns via Scope::atomic_mutate, regex values, named-sub coderefs, folds, pcache, pselect, par_lines, par_walk, par_sed, pwatch, each, four-arg substr, dynamic keys/values/delete/exists, etc.
• Block JIT (strykelang/jit.rs) — Cranelift Block JIT with cached OwnedTargetIsa, tiered after STRYKE_JIT_SUB_INVOKES (default 50) VM invocations. Block JIT validates a CFG, joins typed i64/f64 slots at merges, and compiles hot loops to native code. Disable with --no-jit / STRYKE_NO_JIT=1.
• fusevm bridge (strykelang/fusevm_bridge.rs) — Tier-0 JIT that lowers string-bearing segments (compares, concat, length/ord/hex/oct, index/rindex, crypt, substr/x-repeat with literal strings, sprintf with literal format, length($s) [op] N, mixed-kind chains like length($s) >= $min, etc.) to fusevm's Cranelift block JIT. Strykelang's own block JIT handles pure-int segments; the fusevm bridge covers everything string-bearing through an abstract-stack analyzer with per-slot kind inference. Both try_fusevm_subroutine (sub-call dispatch) and try_fusevm_block_region (map/grep/sort/foreach inline blocks) feed the bridge. Disk-cache safe via STK_VAL_LOAD_CONST (runtime constants resolution that hashes the constant index + helper id, never the per-process heap-pointer bits).
• Per-sub fusevm caches — sub_fusevm_meta: Vec<Option<Option<FusevmSubElig>>> keyed by sub-entry IP (and reused by block-region starts). First call to a hot sub runs 13+ segment_is_* detectors and builds the fusevm chunk via build_chunk; subsequent calls hit OnceCell<Arc<fusevm::Chunk>> for O(1) dispatch. Bench (10M-iter best-of-3, str-bearing micro-benches): chunk cache delivers 9–18% on top of the eligibility cache (cmp −17.6%, length −9.3%, concat −12.9%, sprintf −8.8%, index −11.3%).
• Feature work policy — prefer new VM opcodes in bytecode.rs, lowering in compiler.rs, implementation in vm.rs. Do not add new ExprKind/StmtKind variants for new behavior.
• Parallelism — each parallel block spawns an isolated VM with captured scope; Rayon does work-stealing across all cores.

---

[0x0A] EXAMPLES

examples/ ships 790 top-level .stk programs plus 1349 Rosetta-Code tasks and 173 Exercism solutions — 2.3k working programs in all. Run any of them directly, run the CI sweep with stryke examples/run_all_ci.stk, or run all Exercism solutions with stryke examples/exercism_run_all.stk.

stryke examples/fibonacci.stk
stryke examples/text_processing.stk
stryke examples/parallel_demo.stk
stryke examples/run_all_ci.stk                # validate every example in one pass

Worked examples — long-form

These are full programs (not snippets) that exercise stryke's parallel, AOP, parser, and AI primitives. Each ships with assertions that run on stryke <file> and pass under --no-interop.

# Data + ML
stryke examples/tfidf_search_engine.stk         # TF-IDF inverted index + cosine ranking
stryke examples/gradient_descent_linreg.stk     # mini-batch SGD, pmap_reduce gradient sum
stryke examples/kalman_filter_tracking.stk      # 1D Kalman filter, hand-rolled 2×2 matrix algebra
stryke examples/markov_chain_analysis.stk       # transition matrix + stationary distribution

# Scientific computing — single-file toolkit (124 sections, 23,491 lines, 2,727 hand-crafted die invariants)
stryke --no-interop examples/scientific_compute_no_interop.stk   # vectors, matrices, LU/QR, eigenvalues, root-finding,
                                                                 # quadrature, ODE/PDE, FFT, optimization, statistics,
                                                                 # RNG, graph algos (BFS/DFS/Dijkstra/Tarjan/Kuhn),
                                                                 # string algos (KMP/Z/BM/Aho-Corasick/SA/LCP/Manacher),
                                                                 # DP suites, geometry, number theory, simulation

# Parsers + interpreters
stryke examples/mini_sql_executor.stk           # SELECT/WHERE/ORDER BY/LIMIT recursive-descent
stryke examples/tiny_lisp_parser.stk            # S-expression tokenizer + recursive parser
stryke examples/binary_parser_simulation.stk    # pack/unpack binary protocol with checksum

# Concurrency
stryke examples/pubsub_message_bus.stk          # topic-filtered fan-out over pchannel + spawn
stryke examples/multi_threaded_channels.stk     # bounded pchannel producer/consumer
stryke examples/parallel_pi_monte_carlo.stk     # chunk-parallel Monte Carlo with pmaps + sum
stryke examples/parallel_prime_finder.stk       # pgreps + sort

# Web + protocols
stryke examples/http_router_middleware.stk      # Express-style router, path params, middleware chain
stryke examples/webhook_signature_verifier.stk  # HMAC-SHA256 + replay set + max-age window
stryke examples/network_utilities.stk           # IP sorting / CIDR matching / private-IP check
stryke examples/raft_election_simulator.stk     # discrete-time Raft leader election

# ETL + I/O
stryke examples/csv_etl_parallel.stk            # parse → enrich → ~p> chunk-parallel aggregate
stryke examples/parallel_file_hasher.stk        # ~> glob → head → pmaps sha256 → collect
stryke examples/idiomatic_power_user.stk        # pmaps + hll_count + ai mocked summarize
stryke examples/idiomatic_systems_ops.stk       # AOP `before` audit log + spurt

# AI primitives (mock-mode, no API key required)
STRYKE_AI_MODE=mock-only stryke examples/idiomatic_ai_workflow.stk
STRYKE_AI_MODE=mock-only stryke examples/ai_rag_simple.stk

Idiomatic one-liners + quick demos

# Sets: dedupe + union / intersection
stryke 'val $a = set(1,2,2,3); val $b = set(2,3,4); p len($a | $b), " ", len($a & $b)'

# Threaded pipeline — count primes in 1..100 in parallel
stryke '~> (1:100) pgreps { is_prime _ } collect sort { _ <=> _1 } |> ep'

# File hashes in parallel
stryke '~> glob("*.stk") pmaps { [ _ => sha256(c"#{_}") ] } collect |> ep'

# Sketch algebra — Bloom union
stryke 'val $a = bloom_new(1024); val $b = bloom_new(1024);
        bloom_add($a, $_) for 1:50;
        bloom_add($b, $_) for 30:80;
        p bloom_count($a + $b)'

Rosetta-Code coverage

examples/rosetta/ mirrors the Rosetta Code catalog as fully runnable stryke programs (one file per task) — 1648 tasks at last count. Run a single task with stryke examples/rosetta/<task>.stk or use the project-wide CI in examples/run_all_ci.stk to validate the entire corpus.

Exercism

examples/exercism/ carries 173 idiomatic stryke solutions to public Exercism problems. stryke examples/exercism_run_all.stk runs every solution against its embedded test suite and reports per-track pass counts.

---

[0x0B] BENCHMARKS

stryke vs perl5 vs python3 vs ruby vs julia vs raku vs luajit

bash bench/run_bench_all.sh — stryke vs perl 5.42.2 vs Python 3.14.4 vs Ruby 4.0.2 vs Julia 1.12.6 vs Raku vs LuaJIT on Apple M5 18-core. Mean of 10 hyperfine runs with 3 warmups; includes process startup (not steady-state). Values <1.0x mean stryke is faster.

 stryke benchmark harness (multi-language)
 ──────────────────────────────────────────────
  stryke:  stryke v0.7.7
  perl5:   perl 5.42.2 (darwin-thread-multi-2level)
  python:  Python 3.14.4
  ruby:    ruby 4.0.2 +PRISM [arm64-darwin25]
  julia:   julia 1.12.6
  raku:    Rakudo Star v2026.03
  luajit:  LuaJIT 2.1.1774896198
  cores:   18
  warmup:  3 runs
  measure: hyperfine (min 10 runs)

  bench        stryke ms  perl5 ms  python3 ms  ruby ms  julia ms  raku ms  luajit ms  vs perl5  vs python  vs ruby  vs julia  vs raku  vs luajit
  ---------    ---------  --------  ----------  -------  --------  -------  ---------  --------  ---------  -------  --------  -------  ---------
  startup            3.3       2.3        14.3     23.8      68.3     71.4        1.5     1.43x      0.23x    0.14x     0.05x    0.05x      2.20x
  fib                6.7     184.0        60.1     56.6      76.4    261.3        4.7     0.04x      0.11x    0.12x     0.09x    0.03x      1.43x
  loop               3.2      91.2       191.4     77.8      78.1    159.4        4.3     0.04x      0.02x    0.04x     0.04x    0.02x      0.74x
  string             4.0      10.2        26.8     44.7      83.2    124.2        3.3     0.39x      0.15x    0.09x     0.05x    0.03x      1.21x
  hash               6.8      24.6        25.5     32.6     105.7    143.7        2.0     0.28x      0.27x    0.21x     0.06x    0.05x      3.40x
  array              9.8      24.8        33.2     39.4      88.2    843.9       59.0     0.40x      0.30x    0.25x     0.11x    0.01x      0.17x
  regex             12.6      89.7       264.0    234.3      94.4  25043.8      178.2     0.14x      0.05x    0.05x     0.13x    0.00x      0.07x
  map_grep          13.9      48.8        35.9     48.8      90.5    492.4        3.3     0.28x      0.39x    0.28x     0.15x    0.03x      4.21x

stryke vs perl5 — faster on all 8 benches: fib 27x, loop 29x, regex 7.1x, hash 3.6x, map_grep 3.5x, array 2.5x, string 2.6x, startup 1.4x.

stryke vs python3 — faster on all 8 benches: loop 60x, regex 21x, string 6.7x, fib 9.0x, startup 4.3x, hash 3.8x, array 3.4x, map_grep 2.6x.

stryke vs ruby — faster on all 8 benches: regex 19x, loop 24x, string 11x, fib 8.4x, startup 7.2x, hash 4.8x, array 4.0x, map_grep 3.5x.

stryke vs julia — faster on all 8 benches: loop 24x, startup 21x, string 21x, hash 16x, fib 11x, array 9.0x, regex 7.5x, map_grep 6.5x. Julia timings include LLVM JIT compilation cost — in long-running sessions Julia compiles to native code and would match C on numeric work. These benchmarks measure scripting use cases where startup + single-shot execution matters.

stryke vs raku — faster on all 8 benches by 20-2000x. Raku's regex is 25044ms vs stryke's 12.6ms (1988x). Raku (Perl 6) runs on MoarVM with heavy startup (~70ms+). Raku's strengths are language features (grammars, gradual typing, junctions), not runtime speed.

stryke vs luajit — LuaJIT is the fastest dynamic language runtime ever built (tracing JIT by Mike Pall). stryke beats LuaJIT on 3 of 8 benchmarks: loop (0.74x), array (0.17x), regex (0.07x). Near-parity on string (1.21x) and fib (1.43x). LuaJIT wins on hash (3.4x) and map_grep (4.2x) where its tracing JIT eliminates all dispatch overhead. LuaJIT uses Lua patterns (not PCRE) for the regex bench. stryke offers what LuaJIT cannot: $_, -ne, regex literals, PCRE, parallel primitives (pmap, pmaps, pgrep), streaming iterators, and one-liner ergonomics.

stryke vs perl5 (detailed)

bash bench/run_bench.sh — includes noJIT and perturbation columns for honesty verification. Re-run to get current numbers on your hardware.

Parallel & streaming speedup (100k items, $_ * 2)

  map   (eager, sequential):     0.01s  — inline execution, zero per-item overhead
  maps  (streaming, sequential): 0.11s  — lazy iterator, single interpreter reused
  pmap  (eager, MAX cores):      0.14s  — pre-built interpreter pool, rayon par_iter
  pmaps (streaming, MAX cores):  0.49s  — background worker threads, bounded channel

maps/pmaps are streaming — they return lazy iterators that never materialize the full result list. Use pmaps for pipelines over billions of items where holding all results in memory is impractical, or with take for early termination: range(0, 1e9) |> pmaps { expensive($_) } |> take 10 |> ep.

---

[0x0C] DEVELOPMENT & CI

Pull requests and pushes to main run .github/workflows/ci.yml (Check, Test, Format, Clippy, Doc, Parity, Release Build).

cargo test --lib                # parser smoke, lexer/value/error/scope, interpreter, vm, jit
cargo test --test integration   # tests/suite/* (runtime, readline list context, line-mode stdin, …)
cargo bench --bench jit_compare # JIT vs interpreter on the same bytecode
bash bench/run_bench.sh         # perl5 vs stryke suite (needs hyperfine)
bash bench/run_bench_all.sh     # stryke vs perl5 vs python3 vs ruby vs julia vs raku vs luajit (needs hyperfine)
bash parity/run_parity.sh       # exact stdout/stderr parity vs system perl (20 000+ cases)

• Cargo.lock is committed (CI uses --locked). If your global gitignore strips it, force-add updates: git add -f Cargo.lock.
• Disable JIT: STRYKE_NO_JIT=1 or stryke --no-jit.
• Parity work is tracked in PARITY_ROADMAP.md.

---

[0x0C-test] TEST RUNNER — WORKER POOL ARCHITECTURE

stryke test t/ (or s t t/, or the test() builtin) runs every test_*.stk / t_*.stk file under a directory. The default architecture is a persistent worker pool with fork-on-receive, modeled on cargo-nextest. Process-per-test isolation, no per-test dyld cost, ~5–7× faster than the legacy posix_spawn-per-test path on big corpora.

s t t/                                        # default: worker pool
s t -j 8 t/                                   # 8 worker processes (default = num_cpus)
s t -q t/                                     # quiet — suppress per-file output
s t --no-interop t/                           # forward --no-interop to every test
s t --fork t/                                 # legacy posix_spawn-per-test (slower, fully isolated)
s t --inproc t/                               # single-process VM-per-test (fastest, hermetic only)
test("t/")                                    # builtin form, same default
test("t/", { fork => 1, quiet => 1 })         # builtin opt-out / opts
test_no_interop("t/")                         # builtin variant, --no-interop pinned per worker thread

Topology at peak with -j 18: 1 parent + 18 worker processes + up to 18 grandchildren = up to 37 PIDs, each visible in ps. The 18 OS threads in the parent are not workers themselves — they are pump threads dedicated to one worker process each, just shuttling JSON over stdin/stdout.

parent runner (1 PID)
├── 18 OS pump threads (read JSON results, dispatch from shared crossbeam queue)
├── 18 worker processes (`stryke --test-worker`, persistent, no test code ever runs here)
│   └── on each request: fork() → grandchild
│       └── runs test in fresh VMHelper, writes JSON to saved-stdout fd, _exit
└── result aggregation under `print_lock` (per-block, no line tearing)

Why three modes:

Mode	Per-test cost	Isolation	When to use
--pool (default)	~1 ms (fork-only)	full (own address space)	normal day-to-day
--fork	~9 ms (posix_spawn + dyld + crate static-init)	full	parity / debug
--inproc	~0 ms	shared parent process	hermetic corpora; fastest but tests can corrupt each other

Wire protocol (src/cli_runners.rs) — line-delimited JSON over stdin/stdout pipes:

parent → worker stdin : {"path":"/abs/path/test_foo.stk","no_interop":false,"chdir":"/abs/project_root"}
worker → parent stdout: {"name":"test_foo.stk","passes":18,"fails":0,"failed":false,"detail":null,"stderr":"…"}

Per-test fd dance in the grandchild (so test output never corrupts the wire):

1. dup(1) saves the parent-bound JSON pipe to a private fd.
2. dup2(devnull, 1) — test code's print / say go to /dev/null (would otherwise be parsed by the parent as a malformed JSON response and desync the pool).
3. mkstemp("/tmp/stryke-test-XXXXXX") + unlink + dup2(tmp, 2) — test stderr (the ✓/✗ checkmark lines from test_run) is captured to an anonymous tmp file (auto-deleted on close).
4. Run the test in a fresh VMHelper; pass/fail counters live on the per-VM test_pass_total / test_fail_total atomics.
5. lseek + read the tmp file into the JSON stderr field.
6. write the JSON to the saved stdout fd, _exit(0).

Counter accumulation: the test_run builtin resets per-block counters after printing (so multiple test_run calls in one file work), so embedders read test_pass_total + test_pass_count — totals roll up before the reset; the residual _count is anything that never reached a test_run. Both live on the per-VM VMHelper.

$0: in fork mode the OS sets argv[0] to the test path; in pool mode the runner sets interp.program_name = file_str so tests like slurp(\$0) (used by test_narcissist.stk) keep working.

Multi-root invocations (s t a/t b/t c.stk): the runner groups tests by project_root (parent of t/), chdirs once per group, runs each group's tests in parallel, then moves to the next group. This keeps require "./lib/Foo.stk" working while staying parallel within a single project.

Stragglers: stryke test uses a shared work-stealing queue, so workers naturally rebalance — no static partitioning. The wall-clock floor is the longest-running single test in the slowest worker's tail. Alphabetic ordering tends to put heavy tests late in the queue, so the visible "slowdown at the end" is just the queue draining unevenly. Shortest-job-first ordering would smooth this out but needs a per-test profile pass first.

---

[0x0D] STANDALONE BINARIES (stryke build)

Compile any Perl script to a single self-contained native executable. The output is a copy of the stryke binary with the script source embedded as a zstd-compressed trailer. scp it to any compatible machine and run it — no perl, no stryke, no @INC, no CPAN.

stryke build app.stk                         # → ./app
stryke build app.stk -o /usr/local/bin/app   # explicit output path
./app --any --script --args             # all argv reach the embedded script's @ARGV

What's in the box:

• Parse / compile errors are surfaced at build time, not when users run the binary.
• The embedded script is detected at startup by a 32-byte trailer sniff (~50 µs), then decompressed and executed by the embedded VM. A script with no trailer runs normally as stryke.
• Builds are idempotent: stryke build app.stk -o app followed by stryke --exe app build other.stk -o other strips the previous trailer first, so binaries never stack.
• Unix: the output is marked +x automatically. macOS: unsigned — codesign before distribution if your environment requires it.
• Current AOT runtime sets @INC = ("."); modules outside the embedded script have to be inlined. (require of a local .pm next to the running binary still works.)

Under the hood (strykelang/aot.rs): trailer layout is [zstd payload][u64 compressed_len][u64 uncompressed_len][u32 version][u32 reserved][8B magic b"STRYKEAOT"]. ELF / Mach-O loaders ignore bytes past the mapped segments so the embedded payload is invisible to the OS loader. The b"STRYKEAOT" magic plus version byte lets a future pre-compiled-bytecode payload ship alongside v1 without breaking already-shipped binaries.

# 44 MB binary, no external runtime required:
$ stryke build hello.stk -o hello
stryke build: wrote hello
$ file hello
hello: Mach-O 64-bit executable arm64
$ ./hello alice
hi alice

---

[0x0E] INLINE RUST FFI (rust { ... })

Drop a block of Rust directly into a Perl script. On first run, stryke compiles it to a cdylib (cached at ~/.cache/stryke/ffi/<hash>.{dylib,so}), dlopens it, and registers every exported function as a regular Perl-callable sub.

rust {
    pub extern "C" fn add(a: i64, b: i64) -> i64 { a + b }
    pub extern "C" fn mul3(x: f64, y: f64, z: f64) -> f64 { x * y * z }
    pub extern "C" fn fib(n: i64) -> i64 {
        let (mut a, mut b) = (0i64, 1i64)
        for _ in 0..n { let t = a + b; a = b; b = t; }
        a
    }
}

p add 21, 21         # 42
p mul3 1.5, 2.0, 3.0 # 9
p fib 50             # 12586269025

v1 signature table (parser rejects anything outside this — users write private Rust helpers freely, only exported fns matching the table become Perl-callable):

rust signature	perl call
fn() -> i64 / fn(i64, ...) -> i64 (1–4 args)	integer → integer
fn() -> f64 / fn(f64, ...) -> f64 (1–3 args)	float → float
fn(*const c_char) -> i64	string → integer
fn(*const c_char) -> *const c_char	string → string

Requirements: rustc must be on PATH. First-run compile costs ~1 second; subsequent runs hit the cache and pay only dlopen (~10 ms). #[no_mangle] is auto-inserted by the wrapper — you don't need to write it. The body is #![crate_type = "cdylib"] with use std::os::raw::c_char; use std::ffi::{CStr, CString}; already in scope.

How it works (strykelang/rust_sugar.rs, strykelang/rust_ffi.rs): the source-level pre-pass desugars every top-level rust { ... } into a BEGIN { __stryke_rust_compile("<base64 body>", $line); } call. The __stryke_rust_compile builtin hashes the body, compiles via rustc --edition=2021 -O if the cache is cold, libc::dlopens the result, dlsyms each detected signature, and stores the raw symbol + arity/type tag in a process-global registry. Calls from Perl flow through a fallback arm in [crate::builtins::try_builtin] that dispatches on the signature tag via direct function-pointer transmute — no libffi dep, no per-call alloc, no marshalling overhead beyond the StrykeValue::to_int / to_number / to_string calls you'd do for any builtin.

Combine with AOT for zero-friction deployment: stryke build script.stk -o prog bakes the Perl source — which includes the rust { ... } block — into a standalone binary. The FFI compile still happens on first run of ./prog, but the user only needs rustc once, then the ~/.cache/stryke/ffi/ entry is permanent.

Limitations (v1):

• Unix only (macOS + Linux). Windows support is a dlopen-equivalent swap away but isn't wired.
• Signatures beyond the table above are silently ignored (the function still exists in the cdylib, just not Perl-callable).
• Body must be self-contained Rust with std only — no Cargo.toml / external crate deps. If you need regex or similar, vendor the minimal code into the block.
• The cdylib runs with the calling process's privileges. Trust model is equivalent to do FILE.

---

[0x0F] BYTECODE CACHE (rkyv)

stryke stores compiled bytecode in a single rkyv shard at ~/.stryke/scripts.rkyv. The first run of a script parses + compiles + persists into the shard. Every subsequent run mmaps the shard, validates the archived root once, looks up the entry by canonical path in the zero-copy ArchivedHashMap, and skips lex, parse, and compile entirely.

stryke my_app.stk              # cold: parse + compile + write into the shard
stryke my_app.stk              # warm: mmap shard + lookup + dispatch (skips lex/parse/compile)

Cache invalidation: four conditions all evict a stored entry — no stale bytecode is ever served.

Condition	Trigger
Source mtime mismatch	Edit the .stk file → cache miss → recompile
stryke_version mismatch	Cargo version bump in Cargo.toml
Pointer-width mismatch	Cross-build between 32- and 64-bit targets
Binary mtime newer than cached entry	Rebuild stryke (any cargo build advances target/debug/stryke's mtime) → every cached script invalidates automatically. Catches edits to compiler.rs / parser.rs / vm.rs that don't bump CARGO_PKG_VERSION

Built-in inspection:

cacheview()                    # list all cached scripts with stats
cacheview("pattern")           # filter by path pattern
cacheview("--count")           # just the count

cache_stats()                  # returns {count, bytes, path, enabled}
cache_exists("script.stk")     # 1 if cached, 0 if not
cache_clear()                  # wipe the cache

Example output:

$ stryke -e 'cacheview()'
stryke bytecode cache
  path: ~/.stryke/scripts.rkyv
  scripts: 103 (612.45 KB)

PATH                                                      PROG KB    BC KB
/Users/me/project/lib/heavy_module.stk                       8.57     19.48
/Users/me/project/bin/main.stk                               2.45      5.84
...

Tuning:

• STRYKE_CACHE=0 — disable caching entirely
• Cache is enabled by default for file-based scripts
• Bypassed for -e / -E one-liners (overhead > benefit for tiny scripts)
• Bypassed for -n / -p / --lint / --check / --ast / --fmt / --profile modes

Format: rkyv-archived ScriptShard { header, entries: HashMap<path, ScriptEntry> }. Entries hold per-script (mtime_secs, mtime_nsecs, binary_mtime_at_cache, cached_at_secs, program_blob, chunk_blob). Inner blobs use bincode for now (StrykeValue's Arc-shared graph isn't trivially zero-copy archivable yet — phase 2 will derive Archive directly on Chunk / Program for full zero-copy load). Writes go through flock on scripts.rkyv.lock and atomic rename of a tmp file.

Aligned with zshrs: same rkyv shard pattern (zshrs/src/daemon/shard.rs) — mmap + check_archived_root + zero-copy ArchivedHashMap lookup. zshrs uses per-source-tree shards with a daemon; stryke uses a single global shard since scripts are individually invoked.

Migration rationale: see docs/CACHE_RKYV_MIGRATION.md for the full story — measured 11x speedup on the per-process workload (s test t), tradeoffs, and what's deferred to phase 2.

---

[0x10] DISTRIBUTED pmap_on / ~d> OVER SSH (cluster)

Distribute a pmap-style fan-out across many machines via SSH. The dispatcher spawns one persistent stryke --remote-worker process per slot, performs a HELLO + SESSION_INIT handshake once per slot, then streams JOB frames over the same stdin/stdout. Pairs perfectly with stryke build: ship one binary to N hosts, fan the workload across them.

# Build the worker pool. Each spec maps to one or more `ssh HOST STRYKE --remote-worker` lanes.
val $cluster = cluster([
    "build1:8",                          # 8 slots on build1, default `stryke` from PATH
    "alice@build2:16",                   # 16 slots, ssh as alice
    "build3:4:/usr/local/bin/stryke",        # 4 slots, custom remote stryke path
    { host => "data1", slots => 12, stryke => "/opt/stryke" },  # hashref form
    { timeout => 30, retries => 2, connect_timeout => 5 },  # trailing tunables
])

val @hashes = @big_files |> pmap_on $cluster { slurp_raw |> sha_256) }

# pflat_map_on for one-to-many mapping
val @lines = @log_paths |> pflat_map_on $cluster { split /\n/, slurp }

Distributed thread macro ~d>

~d> is ~p> (the parallel-chunk thread-first macro) but with each chunk shipped to a cluster worker instead of a local rayon thread. Same chunk-block surface — stages operate on @_ (the chunk's elements) and the results merge in source order via the existing pmap_on dispatcher (one persistent ssh process per slot, JOB frames over a shared work queue, per-job retry budget).

val $cluster = cluster("build1:8", "build2:16")

# Distributed equivalent of `~p> @big_files map { sha_256(slurp_raw($_)) }`:
val @hashes = ~d> on $cluster @big_files map { sha_256(slurp_raw($_)) }

# Source-order is preserved even though chunks finish out of order on remote
# workers — the dispatcher tracks per-chunk seq numbers and merges by index.
val @doubled = ~d> on $cluster 1:1_000_000 map { $_ * 2 }
say "$doubled[0] .. $doubled[-1]"        # 2 .. 2000000

The on $cluster operand is required — no implicit default cluster in v1. Trailing ||> / |then| boundary marker switches back to a sequential ~> continuation operating on the auto-merged result, identical to ~p>'s split-boundary semantics:

~d> on $cluster @urls map { fetch($_) } ||> uniq sort

For tests / debugging without an SSH host, STRYKE_CLUSTER_LOCAL_BIN=/path/to/stryke makes the cluster dispatcher spawn the worker locally instead of going through ssh. Slot host fields are ignored when this is set; useful for CI fixtures and single-machine end-to-end smoke tests.

Cluster syntax

Each list element to cluster([...]) is one of:

Form	Meaning
"host"	One slot on host, remote stryke from $PATH
"host:N"	N slots on host
"host:N:/path/to/stryke"	N slots, custom remote stryke binary
"user@host:N"	ssh user override (kept verbatim, passed through to ssh)
{ host => "...", slots => N, stryke => "..." }	Hashref form with explicit fields
trailing { timeout => SECS, retries => N, connect_timeout => SECS }	Cluster-wide tunables (must be the last argument; consumed only when all keys are tunable names)

Tunables (defaults shown):

Key	Default	Meaning
timeout (alias job_timeout)	60	Per-job wall-clock budget in seconds. Slots that exceed this are killed and the job is re-enqueued.
retries	2	Retries per job on top of the initial attempt. retries=2 → up to 3 total tries.
connect_timeout	10	ssh -o ConnectTimeout=N for the initial handshake.

Architecture

main thread                       ┌── slot 0 (ssh build1) ────┐
┌──────────────────┐              │  worker thread + ssh proc  │
│ enqueue all jobs ├──► work_tx ─►│  HELLO + SESSION_INIT once │
│ collect results  │              │  loop: take JOB from queue │
└──────────────────┘              │        send + read         │
        ▲                         │        push to results     │
        │                         └────────────────────────────┘
        │                         ┌── slot 1 (ssh build1) ────┐
        │                         │  worker thread + ssh proc  │
        │                         └────────────────────────────┘
        │                         ┌── slot 2 (ssh build2) ────┐
        │                         │  ...                       │
        │                         └────────────────────────────┘
        │                                    │
        └────────── result_rx ───────────────┘

Each slot runs in its own thread and pulls JOB messages from a shared crossbeam channel. Work-stealing emerges naturally — fast slots drain the queue faster, slow slots take fewer jobs. No round-robin assignment, which was the basic v1 implementation's biggest performance bug (fast hosts sat idle while slow hosts queued). The Interpreter on each remote worker is reused across jobs so package state, sub registrations, and module loads survive between items.

Wire protocol (v2)

Every message is [u64 LE length][u8 kind][bincode payload]. The single-byte kind discriminator lets future revisions extend the protocol without breaking older workers — an unknown kind is a hard error so version skew is loud. See strykelang/remote_wire.rs.

dispatcher                    worker
    │                            │
    │── HELLO ─────────────────►│   (proto version, build id)
    │◄───────────── HELLO_ACK ──│   (worker stryke version, hostname)
    │── SESSION_INIT ──────────►│   (subs prelude, block source, captured lexicals)
    │◄────────── SESSION_ACK ───│   (or ERROR)
    │── JOB(seq=0) ────────────►│   (item)
    │◄────────── JOB_RESP(0) ───│
    │── JOB(seq=1) ────────────►│
    │◄────────── JOB_RESP(1) ───│
    │           ...             │
    │── SHUTDOWN ──────────────►│
    │                            └─ exit 0

The basic v1 protocol shipped the entire subs prelude on every job and spawned a fresh ssh process per item. For a 10k-item map across 8 hosts that's 10 000 ssh handshakes (~50–200 ms each) + 10 000 copies of the subs prelude over the wire — minutes of overhead before any work runs. The v2 persistent session amortizes the handshake across the whole map and ships the prelude once.

Fault tolerance

When a slot's read or write fails (ssh died, network blip, remote crash, per-job timeout), the worker thread re-enqueues the in-flight job to the shared queue with attempts++ and exits. Other living slots pick the job up. A job is permanently failed when its attempt count reaches cluster.max_attempts. The whole map fails only when every slot is dead or every queued job has exhausted its retry budget.

stryke --remote-worker

The worker subprocess. Reads a HELLO frame from stdin, parses subs prelude + block source from SESSION_INIT exactly once, then handles JOB frames in a loop until SHUTDOWN or stdin EOF. Started by the dispatcher via ssh HOST FO_PATH --remote-worker. Also reachable directly for local testing:

echo "..." | stryke --remote-worker      # reads framed wire protocol from stdin
stryke --remote-worker-v1                # legacy one-shot session for compat tests

Limitations (v1)

• Unix only — hardcoded ssh, hardcoded POSIX dlopen path. Windows would need a similar shim.
• JSON-marshalled values — serde_json round-trip loses bigints, blessed refs, and other heap-only StrykeValue payloads. The supported types are: undef, bool, i64, f64, string, array, hash. Anything outside that returns an error from pmap_on.
• varsync / atomic capture is rejected — shared state across remote workers can't honour the cross-process mutex semantics in v1. Use the result list and aggregate locally.
• No streaming results — the dispatcher buffers the full result vector before returning. For huge fan-outs this is the next thing to fix (likely via pchannel integration).
• No SSH connection pool across calls — each pmap_on invocation builds fresh sessions. Subsequent pmap_on calls in the same script reconnect from scratch.

---

[0x08c] --static MODE

stryke's type system is opt-in by default — typed val $x : Int, typed struct/class fields, and typed fn parameters are checked at runtime when present, ignored when absent. --static makes typing mandatory for the whole program, modelled on Kotlin: every function parameter, every return type, and every variable declaration must carry a type (or, for variables, an initializer the type can be inferred from). Statically-known mismatches abort before the program runs; everything else is enforced at runtime by the same check_value path that backs opt-in typing.

--static applies to the main program only — imported .pm modules keep their own (possibly untyped) contract. It also disables the bytecode cache so a non-static cached chunk can never bypass the typing pass.

Type syntax (Kotlin angle brackets)

fn add($x: Int, $y: Int): Int { $x + $y }      # typed params + return type
var $n: Int = 5                                # typed scalar
var $m = 5                                      # OK: Int inferred from initializer
val @names: List<Str> = ("ann", "bob")         # element-typed array
var %ages: Map<Str, Int> = (ann => 30)         # key/value-typed hash
var %nested: Map<Str, List<Int>> = ()          # generics nest

struct Point { x => Int, y => Int }
var $p: Point = Point(x=>1, y=>2)              # struct-typed scalar
var @path: List<Point> = (Point(x=>0, y=>0))   # list of structs/classes
var $seen: Set<Int> = set(1, 2, 3)             # element-typed set
var $pq: Heap<Str> = heap { $a cmp $b }        # element-typed priority queue
var $q: Deque<Int> = deque()                   # element-typed deque

• List<T> is the canonical element-typed array (Array<T> is an alias). Bare Array carries no element check.
• Map<K, V> is the canonical key/value-typed hash (Hash<K, V> is an alias); both require exactly two type arguments. Bare Hash carries no key/value check.
• Value-type containers Set<T>, Heap<T>, Deque<T> are scalar-held (the value is a set / priority queue / deque) and check every member against T.
• Nominal element types — any struct, class, or enum name is a type: Point, List<Point>, Map<Str, Dog>, Set<Color>. A List<Point> rejects a non-Point element; a var $p: Point rejects a non-Point value.
• Return types use the Kotlin colon: fn name(...): Type { }. A capitalized type name after the closing paren is a return type; a lowercase identifier (: lvalue) stays an attribute.

Enforcement

Construct	Without --static	With --static
fn f($x) (untyped param)	allowed	rejected — param needs a type
fn f($x: Int) (no return type)	allowed	rejected — return type required
var $x (no type, no init)	allowed	rejected — type or initializer required
var $x = 5	Any	OK — Int inferred
val $y: Int = "hi"	runtime type error	rejected at parse time
push @a, 5 on List<Str>	—	type error (runtime, both modes once @a is typed)

Typed containers are enforced whether or not --static is on, at every write: declaration, whole-value reassignment, push, and element store ($a[i] = ... / $h{k} = ...). --static adds the requirement that everything be typed plus the parse-time rejection of statically-known mismatches.

Declaring a collection with val (instead of var) makes it immutable: push/element-store on a val @/%, and the in-place mutating methods (push_back/push_front/pop_back/pop_front/push/pop) on a val Heap/Deque, are rejected. Use var for a collection you intend to mutate.

stryke --static script.stk        # whole program must be statically typed
stryke script.stk                 # opt-in typing only (default)

Demo: examples/static_typing_demo.stk — return types, List<T> / Map<K, V> containers, nested generics, every-write enforcement, and inference. Runs under both plain stryke and stryke --static.

---

[0x10a] INFRASTRUCTURE LOAD TESTING

"The hottest language ever created. Literally."

stryke is a server farms first language — the first programming language designed from the ground up for distributed infrastructure load testing. Not HTTP load testing. Not API benchmarks. Bare metal heat.

Stress Testing Builtins

All stress functions pin ALL cores to 100% TDP simultaneously:

stress_cpu(10)           # 10 seconds, SHA256 across ALL cores
stress_mem(1e9)          # 1GB allocated + touched across ALL cores
stress_io("/tmp", 100)   # parallel file I/O across ALL cores
stress_test(60)          # combined CPU + memory + IO stress
heat(60)                 # 🔥 maximum thermal assault

The heat Function

The hottest function in any programming language:

heat(60)

Output:

🔥 HEAT: Pinning MAX cores to 100% TDP for 60s (Ctrl-C to stop early)
🔥 HEAT: 3,116,320,000 hashes in 60.00s (51.9M/s)

Measured Performance (Apple M3 Max)

Function	Result	CPU Usage
stress_cpu(3)	154M hashes	1117% (all cores)
stress_mem(1e9)	1GB touched	452% (parallel)
heat(60)	3.1B hashes	1800% (max TDP)

What stryke Tests

This isn't application performance testing. This is infrastructure validation:

Layer	What You Test
Cooling	Can CRAC units handle sustained full load?
Power	PDU rated for 100% simultaneous draw?
UPS/Generator	Backup power actually works?
Hardware	Which blade has the failing fan?
Ops	Does the NOC notice? How fast?

Distributed Load Testing

Combine with cluster + pmap_on for fleet-wide stress:

val $c = cluster(["node1:16", "node2:16", "node3:16"])

# Pin 48 cores across 3 servers for 60 seconds
1:48 |> pmap_on $c { heat(60) }

Or use the built-in stress_test with cluster:

val $r = stress_test($c, 60)
p "Total hashes: $r->{cpu_hashes}"
p "Workers: $r->{workers}"

Use Cases

• BCP/DR exercises — stress primary datacenter, validate failover
• Capacity planning — prove infrastructure handles peak load
• Burn-in testing — validate new hardware before production
• Cooling validation — find thermal limits before summer hits
• Compliance — demonstrate resilience for SOC 2, PCI DSS, FedRAMP

---

[0x10b] AGENT/CONTROLLER ARCHITECTURE

stryke includes a complete distributed load testing system with persistent agents and interactive control.

Controller (Master REPL)

stryke controller                    # listen on 0.0.0.0:9999
stryke controller --port 8888        # custom port
stryke controller --bind 10.0.0.1    # specific interface

Commands:

Command	Description
status	List connected agents with cores, memory, state
fire [SECS]	Start stress test on all agents (default: 10s)
eval CODE	Run arbitrary stryke source on every agent in parallel. Each agent parses & executes against its own persistent VMHelper, so sub definitions and $main::name globals carry across calls (lexical my/our are per-frame, like a Perl -de0 session). Controller fans out via a two-pass loop — pass 1 writes all EVAL frames (kernel sends, no waiting), pass 2 collects each EvalResult in sorted name order. Wall time = max(per-agent latency), not sum. Output: [agent/ok|ERR] <result> per agent, each output line tagged. 30s read timeout per agent.
@CODE	Shorthand for eval CODE. Any line in the REPL starting with @ ships the rest as stryke source to every agent — @1+1, @sub bump { ... } bump(), @$main::counter += 5 all work. Saves four keystrokes vs the explicit verb.
terminate	Stop stress test immediately
shutdown	Disconnect all agents and exit

eval example session:

stryke controller v0.17.16
> status
node-01    16   64GB         idle      120s
node-02    16   64GB         idle      118s
> @sub greet { "hello from " . $main::ENV{HOSTNAME} } greet()
[node-01/ok] hello from node-01
[node-02/ok] hello from node-02
> @$main::counter = 0
[node-01/ok] 0
[node-02/ok] 0
> @$main::counter += 5; $main::counter
[node-01/ok] 5
[node-02/ok] 5
> @p "a"; p "b"; p "c"; 0
[node-01/ok] a
[node-01/ok] b
[node-01/ok] c
[node-01/ok] 0
[node-02/ok] a
[node-02/ok] b
[node-02/ok] c
[node-02/ok] 0
> @die "boom"
[node-01/ERR] boom at -e line 1
[node-02/ERR] boom at -e line 1
> eval $main::counter             # explicit verb form still works
[node-01/ok] 5
[node-02/ok] 5

Wire-level: a new pair of frame kinds is added to the agent protocol — EVAL (controller → agent, payload = bincode EvalCommand { code }) and EVAL_RESULT (agent → controller, payload = EvalResult { ok, output }). AGENT_PROTO_VERSION is bumped to 2 so a v1 agent refuses the handshake against a v2 controller rather than silently hanging on an unrecognised frame kind.

Builtins: controller(...) and agent(...) — go into either mode from a script

The CLI subcommands stryke controller and stryke agent are mirrored by the controller and agent builtins so any .stk script can drop itself into either mode without going through the CLI dispatch. Both block (the controller on its REPL stdin loop, the agent on its frame loop) and return the exit code as an integer (0 clean shutdown, 1 bind / connect / handshake failure).

# examples/agent_become.stk — script equivalent of `stryke agent`
val $addr = $ENV{CONTROLLER_ADDR} // "localhost:9999"
val $name = $ENV{AGENT_NAME}      // $ENV{HOSTNAME} // "anonymous"
exit agent($addr, $name)

# examples/controller_with_local_agent.stk — controller + local agent in one process
spawn {
    sleep 1                          # let controller bind
    agent("localhost:9999", "local-worker")
}
exit controller("127.0.0.1", 9999)   # blocks on REPL

Signatures:

Builtin	Args	Defaults
controller(bind?, port?)	bind: bind address, port: TCP port	bind="0.0.0.0", port=9999
agent(addr?, name?)	addr: host or host:port, name: display label	addr="localhost:9999", name=hostname

Wrap either in spawn { ... } to run in the background while the calling script continues. Useful for in-process integration tests and local REPL development against a real agent without bringing up a second machine.

[0x10c] SCRIPTABLE DISTRIBUTED COMPUTE — congregation / pray / annex

The controller/agent REPL at [0x10b] is for human-typed interactive load testing. Scripts that need programmatic distributed compute — scatter work to N workers, gather results, manage worker state — use the 28-verb religious-themed API layered on top of the same TCP+bincode controller infrastructure. Each verb is a stryke builtin; the whole pipeline lives in a .stk file with no REPL, no manual orchestration, no external infrastructure.

Minimum-viable use

val @workers = congregation(4);                # fork 4 agents locally, wait for them to register
val $div     = pray "compute()", @workers;     # scatter EVAL frames in parallel, return divination
val %results = annex $div;                     # block-and-gather hash keyed by session-id
excommunicate(@workers);                      # clean shutdown

pray accepts a string OR a coderef (closure body is deparsed and shipped — closure captures not supported in v1):

val $div = pray fn { 2 + _ }, @workers;         # coderef form
val $div = pray "2 + 3", @workers;              # string form

Full verb taxonomy

Verb	Side	Effect
Lifecycle
congregation($n)	master	fork N stryke agent children, auto-register them, return handle array
anoint($n)	master	like congregation but doesn't take over as current controller (multi-cong scripts)
ordain([$name, [$bind, [$port]]])	master	bare controller (no agents), register name in cathedral for remote profess
muster([$controller_id])	master	enumerate currently-connected agent handles
welcome($n [, $timeout_ms])	master	block until N agents have joined
excommunicate(@handles)	master	SHUTDOWN frame to subset, drop from roster
bow()	slave	enter agent receive loop (alias for agent())
profess($name)	slave	look up congregation in cathedral, connect as agent
apostatize($name)	local	unregister congregation from cathedral
cathedral()	local	enumerate registered congregation names
Scatter / gather
pray($code, @handles)	master	scatter, return divination id (closure or string)
annex($div [, $timeout_ms])	master	block-and-gather, consume divination, return hash
harvest($code, @handles [, $timeout_ms])	master	fused pray + annex one-shot
chant($code, @handles)	master	continuous rescatter — fires at current AND future joiners; returns chant_id
amen($id)	master	release pending divination OR stop active chant
pilgrimage($code, @handles [, $timeout_ms])	master	scatter+gather barrier; returns 1 if all rendezvous, 0 otherwise
State inspection
lick(@handles)	master	non-destructive snapshot of every worker's %soul via to_json(\%soul)
peruse(@handles)	master	deep %soul walk (Tier 3 alias of lick; god-style traversal in Tier 5)
interrogate($pid_or_handles)	master	polymorphic dump — $pid (OS process via sysinfo) OR @handles (agent VM state)
State mutation
bestow(\%hash, @handles)	master	push hash to each worker's %gift via JSON round-trip
smite(@handles)	master	reset workers' %soul and %gift without disconnecting
recant(@keys)	slave	partial self-erasure of own %soul
Persistence
enshrine(\%hash, $path)	local	write hash to disk as JSON
exhume($path)	local	read enshrined JSON back as hash
smother(\%hash)	local	securely zero a local hash (overwrite + clear)
martyr($path)	slave	exit(0) after caller has enshrined
resurrect($enshrine_path)	master	exhume + anoint(1) + bestow → new agent with restored state
Security
cloister($token)	master	toggle :cloistered mode — agents must send STRYKE_AGENT_TOKEN matching $token
divine($handler)	slave	register closure handler for incoming petitions (Tier 5 wires dispatch through it)

Worker-side %soul convention

Workers maintain our %soul as their externally-visible state. lick/peruse/annex serialize it; smite clears it; bestow populates a peer hash our %gift for pushed data:

# Slave script (run on each compute node):
profess "renderfarm";
bow {                                          # enter receive loop
    our %soul;
    our %gift;
    our $frames_rendered = 0;
};
# Master can later: lick(@workers) sees the soul state; smite(@workers) resets it

:cloistered ACL

Open by default. To restrict membership to anointed-only workers, the master calls cloister($token) and workers set STRYKE_AGENT_TOKEN matching that token before calling agent/bow/profess. Wire-level: a new AGENT_AUTH frame (frame_kind::AGENT_AUTH = 0x1B) is sent post-HELLO when the env var is set; cloistered controllers reject any agent without a valid token within 500 ms.

Continuous rescatter via chant

chant is the fire-and-forget cousin of pray — it registers an ongoing prayer that fires at every current agent AND at every new agent that joins later (via the accept-loop hook). Use for state distribution (bestow-like config push to current + future workers):

val $vigil = chant("our %config = (max_depth => 8); 'ok'", @workers);
# ... new workers can join via profess(); they auto-receive the chant on join ...
amen($vigil);                                  # stop the rescatter

Cathedral (in-process v1)

ordain($name, ...) registers name → master_endpoint in an in-process registry. profess($name) resolves the endpoint and connects. Tier 4 ships in-process only — Tier 5 promotes the cathedral to a standalone stryked daemon for cross-host name resolution.

# Master:
ordain "renderfarm", "0.0.0.0", 9999;
welcome 4, 30_000;                             # wait for 4 slaves
val %frames = harvest "render_frame()", muster();

# Slave (in any process, including a forked child):
profess "renderfarm";                          # blocks in agent loop

Live demos

14 demos under examples/, all CI-safe (loopback fork only, no network) and clean under --no-interop:

Demo	What it shows
distributed_congregation.stk	Tier 0 minimum-viable scatter-gather (2 workers, 1 prayer)
congregation_100x_scale.stk	Large-fleet scatter-gather; tested clean at 100 and 250 workers
distributed_prime_sieve.stk	Real CPU work — shard 0..10000 across N workers, verify π(10000)=1229
distributed_wordcount.stk	MapReduce shape — every worker counts words, master verifies agreement
distributed_log_aggregation.stk	Telemetry pattern — per-event JSON counts, master sums fleet-wide
harvest_oneshot.stk	The ergonomic shape (harvest $code, @workers = pray + annex fused)
bestow_then_lick.stk	Master pushes config via bestow, reads worker state back via lick
pilgrimage_barrier.stk	3-stage BSP barrier across 4 workers
congregation_chant_late_joiners.stk	chant/amen continuous-rescatter lifecycle
smite_state_reset.stk	Reset workers' %soul without disconnecting (vs excommunicate)
enshrine_exhume_roundtrip.stk	Persist hash to disk as JSON, exhume back, verify identity
cloistered_acl_demo.stk	:cloistered ACL + cathedral inspection + apostatize cleanup
interrogate_pids.stk	Polymorphic interrogate($pid) (OS) vs interrogate(@handles) (agent VM state)
multi_congregation.stk	anoint spawns secondary congregation alongside primary

Sample run (the original Tier 0 demo):

$ stryke examples/distributed_congregation.stk 2 "5 * 8"
── distributed_congregation: N=2 code=5 * 8 ──
spawned 2 workers: 1,2
divination id: 1
  worker 1 → 40
  worker 2 → 40
✓ all 2 workers returned: 40

100-worker scale demo:

$ stryke examples/congregation_100x_scale.stk 100
── 100x scale: spawning 100 worker processes ──
spawned 100 / 100 workers (fork + connect + register)
harvested 100 / 100 replies
  matched:  100 / 100 replies
  diverged: 0
✓ every spawned worker returned the correct answer
excommunicated 100 workers

Project-bar position

Scriptable in-language single-keyword scatter-gather + distributed state harvest + secure-erase as a language primitive does not exist anywhere as of 2026-05-27. MPI is a C library requiring mpirun; Erlang has process groups but no destructive harvest or peek/commit pair; Spark/Hadoop require cluster bootstrap and JVM; nothing in shell-language space comes close. The lick/annex/smother triple and the chant/amen continuous-rescatter pair are genuinely empty territory — world-first language-keyword primitives, not yet-another-rewrite of an existing framework.

See docs/killer-features-brainstorm.md "Scriptable Master/Slave" for the full design rationale, shipped/deferred status, and the two stryke language bugs fixed during this work (\%our-hash ref deref + our %hash cross-EVAL persistence).

Builtins: mark(...) / provenance(...) / unmark(...) — value lineage as a first-class feature

World-first for scripting languages: no major scripting language (Perl, Python, Ruby, JavaScript, Lua, PHP) ships automatic value-lineage tracking as a first-class builtin. Closest analogs are research dataflow languages (LIO, Adapton) which surface lineage as a type-system feature, never as a plain provenance($x) call. Stryke ships the trio as a normal dispatch builtin with O(1) HashMap-keyed lookup and zero overhead until any value is marked.

val $config = mark({ host => "prod", retries => 3 })       # tag the Arc
val $p      = provenance($config)
# { origin => "HASH entries=2", origin_line => 1, ops => [] }
unmark($config)                                            # reclaim ledger entry

Builtin	Signature	Returns
mark($val)	tag a heap value's Arc; idempotent	$val unchanged (pipeline-friendly)
provenance($val)	look up $val's lineage	{ origin, origin_line, ops => [...] } or undef
unmark($val)	drop the ledger entry	$val unchanged

Cost model: zero when unused. A process-global LEDGER_ACTIVE AtomicBool flips to true on the first mark(...) call. Until then every dispatch's post-hook elides via a single inlined load(Relaxed) branch.

v1 scope (intentional — documented in strykelang/provenance.rs and the LSP hover doc):

• Tracks builtin-call op chains only. User-sub call boundaries aren't recorded yet.
• Keys on heap Arc pointer — two structurally-equal values with different origins have independent lineages; two refs to the SAME Arc share lineage (the aliasing model god already uses).
• Immediates (integers, floats, undef) have no Arc — provenance on them returns undef. Wrap in a hashref if you need to track a scalar.
• String results from builtins (to_json, sha256, base64_*, ...) lose Arc identity in the VM's scalar-return path. Workaround: wrap the string in a one-key hashref so the container's stable Arc carries the lineage. Heap-container results (arrays, hashes, atomics, sets, deques, byte buffers) propagate correctly.
• Ledger entries persist until unmark($val). A v2 sweep based on Arc weak refs is on the roadmap.

Demo: examples/provenance_basics.stk walks through each shape end-to-end.

Builtins: kick(...) / udp_send(...) — TCP knock + UDP multi-shot

Convenience builtins over standard socket calls. Both capabilities exist in every language's stdlib; stryke ships them as bare builtins so service probes / Wake-on-LAN scripts / NAT keepalives don't need a socket import. (punch — full NAT hole-punching with STUN discovery for peer-to-peer UDP between stryke instances behind NAT — lands in a follow-on commit.)

# TCP service-health sweep — 250 ms per probe, returns 1 / 0.
for val $pair (([ "db",  5432 ], [ "redis", 6379 ], [ "api", 8080 ])) {
    printf "  %-6s :%d  %s\n", $pair->[0], $pair->[1],
        kick($pair->[0], $pair->[1], 250) ? "UP" : "down"
}

# Wake-on-LAN magic packet, 3× for reliability.
val @mac_bytes = map { hex($_) } split /:/, "aa:bb:cc:dd:ee:ff"
var @packet = (0xff) x 6
push @packet, @mac_bytes for 1:16
udp_send("255.255.255.255", 9, pack("C*", @packet), 3)

Builtin	Signature	Returns
kick($host, $port [, $timeout_ms])	TCP connect with timeout (default 1000 ms)	1 on success, 0 on any failure
udp_send($host, $port, $payload [, $retries=1, $interval_ms=20])	bind ephemeral, sendto the payload N times	count of successful sends

Both fail soft — bad host, closed port, DNS failure, invalid port all return 0 without raising, so callers can write if (kick(...)) and if (udp_send(...)) idiomatically.

Demo: examples/kick_probe.stk.

Builtins: udp_open / stun / punch / udp_send_to / udp_recv / udp_close — P2P over the open internet

Stryke-to-stryke communication between two hosts behind arbitrary NATs, no infrastructure required. STUN client + UDP hole-punching state machine, both rolled in-tree (~500 lines of Rust, no third-party network crate dependency). Bytecode-VM-direct dispatch — same path as the rest of the builtins.

# Side A (any host, runs first):
val $sock = udp_open()
val $info = stun($sock)                       # → { public_ip, public_port }
printf "Tell peer to: stryke peer.stk %s %d\n",
    $info->{public_ip}, $info->{public_port}
val $first = udp_recv($sock, 60_000)          # wait for peer's bombards
p "Peer connected: $first"
udp_close($sock)

# Side B (peer host):
val $sock = udp_open()
val $r = punch($sock, $peer_ip, $peer_port, { payload => "hello!" })
if ($r->{established}) {
    printf "Connected in %d bombards (%dms)\n", $r->{bombards}, $r->{latency_ms}
    p "Peer replied: $r->{peer_msg}"
}
udp_close($sock)

Builtin	Signature	Returns
udp_open([$bind_host, $bind_port])	bind a UDP socket, register in pool	integer handle (0 on bind failure)
udp_send_to($id, $host, $port, $payload)	send via pool socket	bytes sent (0 on failure)
udp_recv($id [, $timeout_ms=1000])	receive one datagram (payload only)	payload string / bytes / undef on timeout
udp_recv_from($id [, $timeout_ms=1000])	receive one datagram + source address	{ payload, src_ip, src_port } or undef
udp_close($id)	release socket from pool	1 if present, 0 if unknown
stun($id [, $stun_host, $stun_port, $timeout_ms])	query STUN server via socket	{ public_ip, public_port } or undef
stun_classify($id [, $opts])	query multiple STUN servers, detect symmetric NAT	{ nat_type, public_ip, queried, succeeded, observations }
punch($id, $peer_ip, $peer_port [, $opts])	hole-punching state machine	{ established, latency_ms, bombards, peer_msg, peer_addr }

Protocol details: RFC 8489 STUN Binding Request with XOR-MAPPED-ADDRESS parsing (modern form) + MAPPED-ADDRESS fallback (legacy servers). Default STUN server is stun.l.google.com:19302 — Google's public, free, reliable. The same socket handle MUST flow through STUN → punch → application traffic because the NAT mapping is tied to the socket's (local_ip, local_port).

Real-world success rate is ~70-80%. Failures (no language-level workaround):

• Symmetric NATs — some mobile carriers, some corporate firewalls assign a different public port per destination, so the port revealed by STUN ≠ the port the peer's traffic arrives on.
• UDP-blocking firewalls — drop all UDP outright.
• Timing misalignment — if one peer punches 30s before the other, the first peer's NAT mapping may time out before the second peer starts.

For guaranteed delivery use the TURN relay builtins below (turn_allocate, turn_permission, turn_send, turn_recv, turn_refresh) — relay every packet through a public coturn server, works regardless of NAT type.

Demo: examples/p2p_chat.stk — runnable two-process flow.

Builtins: turn_allocate / turn_permission / turn_send / turn_recv / turn_refresh — TURN relay fallback (RFC 8656)

When stun_classify reports symmetric (~20-30% of real-world NAT scenarios — mobile carriers, some corporate firewalls), pure hole-punching cannot succeed. TURN routes every packet through a public relay server you trust, working regardless of NAT type at the cost of one network hop and the bandwidth budget of whoever runs the TURN server.

RFC 8656 client implemented in-tree (~600 lines, no third-party network crate — reuses the existing hmac + sha1 + md-5 from RustCrypto for MESSAGE-INTEGRITY auth). Builds on the STUN binary protocol from nat_punch.rs; TURN messages are STUN-formatted frames with different message types and attributes.

# Side A (the listener):
val $sock = udp_open()
val $alloc = turn_allocate($sock, "turn.example.com", 3478, "alice", "hunter2")
printf "Tell peer my relay is: %s:%d\n", $alloc->{relay_ip}, $alloc->{relay_port}
while (defined(val $msg = turn_recv($sock, 0))) {
    p "from $msg->{peer_ip}:$msg->{peer_port}: $msg->{payload}"
    turn_send($sock, $msg->{peer_ip}, $msg->{peer_port}, "ack")
}

# Side B (the sender):
val $sock = udp_open()
val $alloc = turn_allocate($sock, "turn.example.com", 3478, "bob", "trustno1")
turn_permission($sock, $peer_relay_ip)
turn_send($sock, $peer_relay_ip, $peer_relay_port, "hello via turn!")
val $reply = turn_recv($sock, 5_000)

Builtin	Signature	Returns
turn_allocate($id, $server, $port, $user, $pass [, $timeout_ms])	two-roundtrip auth + Allocate	{ relay_ip, relay_port, lifetime_secs, realm, nonce_len } or undef
turn_permission($id, $peer_ip [, $timeout_ms])	install CreatePermission for peer	1 / 0
turn_send($id, $peer_ip, $peer_port, $payload)	wrap in SEND-INDICATION, ship via relay	bytes sent to TURN server / 0
turn_recv($id [, $timeout_ms])	parse next DATA-INDICATION	{ payload, peer_ip, peer_port } or undef
turn_refresh($id [, $lifetime_secs, $timeout_ms])	extend allocation (or release with lifetime_secs=0)	new lifetime / 0

Protocol scope (RFC 8656 core, deliberately bounded):

• ✅ Allocate with long-term HMAC-SHA1 auth + nonce/realm exchange
• ✅ CreatePermission
• ✅ SendIndication / DataIndication
• ✅ Refresh
• ✅ IPv4 + IPv6 XOR address parsing
• ❌ ChannelBind / ChannelData — bandwidth optimization, not v1
• ❌ TLS/DTLS transport — plain UDP for v1
• ❌ SHA-256 / SHA-384 MESSAGE-INTEGRITY — coturn defaults to SHA1
• ❌ ALTERNATE-SERVER redirect handling

Verified end-to-end via an in-process mock TURN server (8 unit + 3 integration pins) — for real coturn interop you should test against your own server.

Demo: examples/turn_relay_chat.stk — three modes (classify NAT type / listen for messages via relay / send via relay).

Active probes — tcp_probe / tcp_banner / whois_query

Sibling to kick (which returns 1/0) — pick by what you need:

Builtin	Returns	When to use
kick($host, $port)	1 / 0	Fast service-mesh sweep, don't care about latency
tcp_probe($host, $port [, $timeout_ms])	{ alive, latency_ms }	Same probe, but with the RTT measurement
tcp_banner($host, $port [, $timeout_ms, $max_bytes])	{ alive, latency_ms, banner }	Service fingerprint (SSH version, HTTP server header, SMTP greeting)
whois_query($domain [, $server, $timeout_ms])	response string or undef	Domain registration / IP-ownership investigation (RFC 3912)

# SSH version sniff across a fleet.
for val $host (qw(web1 web2 db1)) {
    val $r = tcp_banner($host, 22, 500)
    printf "  %-8s %s\n", $host, \$1 if $r->{alive} && $r->{banner} =~ /^SSH-(\S+)/
}

# Find the authoritative WHOIS server for a TLD, then chase the refer:
val $iana = whois_query("example.com")
val ($registry) = $iana =~ /^refer:\s+(\S+)/m
val $detail = whois_query("example.com", $registry // "whois.verisign-grs.com")

Demo: examples/network_recon.stk chains tcp_probe + tcp_banner via pmap for parallel asset discovery + service fingerprinting.

Builtins teleport / arrive — multi-target SHM IPC

Broadcast a value to N stryke processes via one POSIX shared-memory segment plus per-receiver Unix-domain-socket notification. For big payloads, only one process copies the bytes into kernel memory; each receiver mmaps the same backing pages.

World-first: no other scripting language ships single-call zero-copy(ish) multi-target value teleportation as a primitive. Closest analogs — Python multiprocessing.shared_memory, Perl IPC::Shareable, Ruby mmap — are single-target, require manual segment lifecycle + per-process attach plumbing, and don't bundle a notification primitive. teleport is one call.

Builtin	Signature	Returns
teleport	teleport($val, @receiver_pids [, { hold_ms => 500 }])	count of receivers whose UDS accepted the notify
arrive	arrive([$timeout_ms=5000])	the teleported value (deeply ref-wrapped) or undef on timeout

Surfaces — all four forms are supported:

# Standalone with literal PIDs:
teleport($payload, $pid1, $pid2, $pid3)

# Arrayref of PIDs:
teleport($payload, [@worker_pids])

# Opts hash for hold window:
teleport($payload, @pids, { hold_ms => 1500 })

# Prefix thread macro (data-first signature makes this natural):
~> $payload teleport(@worker_pids)

Wire path: StrykeValue → serde_json::Value → JSON bytes → POSIX SHM segment named /stryke_tp_PID_SEQ→ 40-byte notify over UDS at/tmp/stryke_teleport_PID.sock per receiver. Receivers reverse this. Same lossy-but-portable shape as cluster / dist_thread IPC — scalars, arrays, hashes, and nested combos round-trip; closures and blessed objects don't.

Top-level + nested containers always arrive as refs (matching decode_json's convention) so $msg->{key}->[i]->{inner} Just Works without re-wrapping.

Wire format limits (v1, intentional):

• Receivers must be stryke processes running an arrive() loop. Non-stryke processes or stryke processes without arrive() have no bound UDS and count as unreachable.
• macOS POSIX SHM names cap at 30 chars; /stryke_tp_99999_99 is 19 chars so we have headroom for billions of segments per PID.
• Sender holds the SHM segment alive for hold_ms (default 500), not ack-based. Receivers slower than hold_ms get a stale-name shm_open failure.
• No encryption — receivers are cooperating processes; payload is visible to anything that knows the SHM name.

Fork-safety: the receiver's UDS socket is PID-aware. After fork(), the child rebinds at its own /tmp/stryke_teleport_PID.sock on first arrive() call (the parent's inherited fd is dropped). So fork-then-arrive in the child Just Works without manual reinitialization.

Demo: examples/teleport_broadcast.stk — parent forks N workers, builds a sharded corpus hashref, teleports the whole thing once; each worker reconstructs and processes its slice.

Common pitfalls — friction points worth knowing upfront

These are real bumps I hit while building the NAT-traversal stack — surfacing them so you don't waste time on the same investigations.

Pitfall	Symptom	What to do instead
Postfix for after printf/print	printf "..." for @arr fails with Expected LParen, got ArrayVar(...)	Use explicit block form: for val $x (@arr) { printf "...", $x }
Postfix if after printf	Same parse failure as above	Same fix — wrap in explicit if (...) { ... } block
$tx->clone on pchannel	"Can't call method on non-object"	Multi-producer channels aren't supported; use a different shape (multiple receivers, or one producer fanning to N consumers)
String return → provenance lost	mark({...}); val $j = to_json(...); provenance($j) → undef	Wrap the string in a one-key hashref: mark({ payload => to_json(...) }). VM re-Arcs scalar string returns, breaking ptr-keyed lookup. Document in provenance v1 limits.
grep { Pkg::fn } doesn't auto-bind $_	All elements pass / fail uniformly (predicate sees $path = undef)	Pass explicitly: grep { Pkg::fn($_) } @list
fn($a, @b, @c) slurps	Second @arr param always empty; first @arr contains both	Pass arrayrefs + deref inside: fn ... ($a, $b, $c) { val @b = @$b; val @c = @$c; ... }
mark(\@arr) vs mark([...])	\@arr produces a fresh SCALARREF Arc per access → provenance lookup misses	Use anonymous arrayrefs (mark([10, 20])) or hashrefs — they have stable Arc identity. \@arr operator semantics are subtle
pack "a*" for variable-length string	pack: 'A' and 'a' do not support '*'	Concat: pack("a4 n", $magic, $len) . $payload — stryke's pack is stricter than Perl's
par { Pkg::fn } is chunked, not 1:1	Result count = worker thread count, not input count	Use pmap { Pkg::fn($_) } @list for 1-result-per-input. par runs BLOCK once per chunk with _ = whole chunk list
Tests that bind+drop a port then probe it	Flaky under parallel-test load — another test grabs the freed port between drop and probe	Use port 1 (privileged, never auto-assigned) for "guaranteed closed" probes; or hold the listener for the test duration
Test file reads with CWD-relative paths	Pass when cargo test is run from repo root, NotFound when run from elsewhere (IDE runner, cargo test --manifest-path ...)	Build absolute paths via env!("CARGO_MANIFEST_DIR"). See tests/suite/examples_strict_lint.rs for the pattern

NAT traversal — quick reference

Compact decision sheet. Each rung has a different cost / success / failure profile; pick by stun_classify outcome + your tolerance for infrastructure.

Rungs ranked by cost (low → high) and success rate (network-dependent):

Rung	Cost	Success rate	Fails when	Use when
1. host (direct UDP)	0 (just udp_send_to)	LAN: 100% · public IP: 100% · NAT: 0%	both peers behind any NAT	LAN or one-side-public-IP
2. server-reflexive (punch)	1 STUN RTT + ~5-30 hole-punch bombards	cone NAT: ~95% · symmetric NAT: 0% · UDP-blocked: 0%	symmetric NAT or UDP-blocked firewall	stun_classify returned cone
3. relayed (turn_*)	TURN allocation + 1 hop per packet	~100% (assuming TURN reachable)	TURN server unreachable / unauthenticated	symmetric NAT, UDP-blocked firewall, OR cone-NAT fallback when rung 2 fails

Decision tree (codifies the same logic ice::connect automates):

                  stun_classify($sock)
                          ↓
        ┌─────────────────┼─────────────────┐
        ↓                 ↓                 ↓
     "cone"           "symmetric"        "unknown"
        ↓                 ↓                 ↓
   try host first      skip punch       try host, then
   then punch          go straight      punch, then
                       to turn          turn (probe-best)

Per-step latency (rough, on a normal home connection):

Operation	Typical wall time
udp_open	< 1 ms
stun_classify (3 servers in parallel)	~30-80 ms (slowest STUN RTT + handshake)
stun (single server)	~10-50 ms
punch (cone-to-cone)	100 ms - 5 s (timeout governs)
turn_allocate (incl. 401 + auth retry)	~40-100 ms
turn_send / turn_recv data path	+5-30 ms vs direct (TURN relay hop)

What can go wrong + what to do:

Symptom	Likely cause	Fix
stun_classify returns "unknown"	< 2 STUN servers reached	Check internet, add more servers in $opts->{servers}
punch returns established=0 after timeout	One side behind symmetric NAT, OR misaligned bombard timing	Fall back to TURN; or have peers retry simultaneously
turn_allocate returns undef	Bad credentials, server unreachable, or non-RFC-8656 server	Test with examples/turn_health_check.stk against your server
Data arrives but udp_recv_from returns wrong src_ip	NAT rewrote the source — normal	Use the address as-is; that's the public address the peer's NAT mapped to

Demos showing each shape: p2p_chat.stk (raw primitives) · p2p_chat_v2.stk (ice::connect) · turn_relay_chat.stk (TURN path) · turn_health_check.stk (TURN server probe) · port_scanner.stk (kick+pmap) · network_introspect.stk (reflection-driven builtin discovery)

P2P walkthrough: pick the right rung for YOUR network

Decision tree for stryke-to-stryke P2P over the open internet — what to call, in what order, what each return value means:

val $sock = udp_open()                          # 1. Bind a UDP socket once
                                                #    (used for STUN + the data path)

val $nat = stun_classify($sock)                 # 2. Detect NAT type (~3-server query)
if ($nat->{nat_type} eq "symmetric") {         #    Symmetric → hole-punching can't work
    die "need TURN — see turn_allocate below"  #    (room for a TURN fallback here)
}

val $me = stun($sock)                           # 3. Discover OWN public address
printf "my address: %s:%d — send this to peer\n",
    $me->{public_ip}, $me->{public_port}
# ... exchange addresses with peer via email/paste/IRC/anything ...
val $peer_ip   = "203.0.113.45"
val $peer_port = 51234

val $r = punch($sock, $peer_ip, $peer_port,     # 4. Hole-punch the peer
    { timeout_ms => 5000 })
if (!$r->{established}) {
    die "punch failed — peer NAT too restrictive; need TURN"
}
printf "connected: %s\n", $r->{peer_msg}      # 5. Bidirectional flow established

# Now send/recv as normal UDP via the same socket:
udp_send_to($sock, $peer_ip, $peer_port, "hello")
val $reply = udp_recv_from($sock, 5000)        # → { payload, src_ip, src_port }
p "peer said: $reply->{payload}"
udp_close($sock)

If stun_classify returns	Then you can	Otherwise
cone	use punch directly (~70-80% success)	—
symmetric	skip punch — it WILL fail	use turn_allocate + turn_send / turn_recv
unknown	try punch first, fall back to TURN on failure	—

The orchestrator at examples/ice_orchestrator.stk automates this whole ladder — see the next subsection. The two runnable demos pair like this:

• examples/p2p_chat.stk — uses raw primitives (udp_open, stun, punch, udp_send_to, udp_recv_from). Read this first to understand what's happening on the wire.
• examples/p2p_chat_v2.stk — uses ice::connect from the orchestrator. Copy this into your own code.

ICE-lite: orchestrating direct → punch → relay in stryke source

The v1.3 primitives (udp_open / stun / stun_classify / punch / turn_*) compose into a complete NAT-traversal orchestrator. The orchestration algorithm — which transport to try first, when to fall back — is intentionally stryke source rather than a Rust builtin: the wire-protocol code is already builtins, and keeping the ladder logic as user-editable source means you can read it, adapt it, and inline custom rules without recompiling.

examples/ice_orchestrator.stk implements the three-rung ladder per RFC 8445 §5 (the connectivity-check core):

Rung	Transport	Works when
1. host	udp_send_to direct to peer's host:port	peer has no NAT (public IP / same LAN)
2. server-reflexive	punch using STUN-discovered address	both sides behind cone NAT (~70% of pairs)
3. relayed	turn_* via TURN server	always (assuming TURN is reachable)

require "examples/ice_orchestrator.stk"

val $conn = ice::connect({
    peer_host_addr   => "192.0.2.50:9000",      # try direct first
    peer_srflx_addr  => "203.0.113.45:51234",   # peer's STUN-discovered
    peer_relay_addr  => "198.51.100.99:49000",  # peer's TURN allocation
    turn_server      => "turn.example.com:3478",
    turn_user        => "alice",
    turn_pass        => "hunter2",
    timeout_ms       => 3000,
})
die "no transport: $conn->{reason}" unless $conn->{ok}
# $conn->{method} is "direct" | "punch" | "relay"; $conn->{socket} is the
# bound socket to use for subsequent send/recv (udp_send_to OR turn_send
# per method).

ice::gather_candidates({ turn_server, turn_user, turn_pass }) gathers the local side's three candidates so you can publish them via signaling (email, paste, IRC, anything) for the peer to use.

What this ICE-lite does NOT cover (RFC 8445 proper):

• Full priority math + role tie-breakers (8445 §5.1.2 / §5.2)
• Regular vs aggressive nomination (§8)
• Conflict resolution when both peers think they're the controller (§7.1.3.3)
• TCP candidates (RFC 6544)
• Trickle ICE (RFC 8838)

These are useful in WebRTC where the orchestrator runs autonomously; for stryke's typical "two peers with manual signaling" case the three-rung ladder + first-success wins is enough and stays in ~200 lines of inspectable code.

Agent (Worker Daemon)

stryke agent                              # use config file
stryke agent --controller 10.0.0.1        # connect to specific host
stryke agent --port 9999                  # specific port

Config file: ~/.config/stryke/agent.toml

[controller]
host = "controller.example.com"
port = 9999

[limits]
max_temp = 85       # auto-terminate if CPU temp exceeds
max_duration = 3600 # max seconds per session

[agent]
name = "node-01"    # optional, defaults to hostname

Example Session

$ stryke controller
stryke controller listening on 0.0.0.0:9999
[agent connected] node-01 (cores=64, session=1)
[agent connected] node-02 (cores=64, session=2)
[agent connected] node-03 (cores=64, session=3)

controller> status
AGENT                 CORES     MEMORY        STATE       UPTIME
------------------------------------------------------------
node-01                  64       256GB         idle         42s
node-02                  64       256GB         idle         38s
node-03                  64       256GB         idle         35s

Total: 3 agents, 192 cores, 0 firing

controller> fire 60
[fire] 3 agents, duration=60s

controller> terminate
[terminate] 3 agents

Wire Protocol

Framed binary protocol over TCP:

CONTROLLER                    AGENT
    │                           │
    │◄──── AGENT_HELLO ─────────│  (hostname, cores, memory)
    │───── AGENT_HELLO_ACK ────►│  (session_id)
    │                           │
    │───── FIRE ───────────────►│  (workload, duration)
    │◄──── METRICS ─────────────│  (cpu%, hashes/sec)
    │───── TERMINATE ──────────►│
    │◄──── TERM_ACK ────────────│  (final stats)

Deployment

Single binary, zero dependencies:

# Build self-contained agent binary
stryke build agent.stk -o stryke-agent

# Deploy to any Linux server
scp stryke-agent node1:/usr/local/bin/
ssh node1 'stryke-agent --controller controller:9999'

Kubernetes DaemonSet:

apiVersion: apps/v1
kind: DaemonSet
metadata:
  name: stryke-agent
spec:
  template:
    spec:
      containers:
      - name: agent
        image: ghcr.io/menketechnologies/stryke:latest
        args: ["agent", "--controller", "stryke-controller:9999"]

Why This Matters

Other Tools	stryke
External load generators	Agents inside cluster
Config files, YAML, XML	fire 60 — two words
Batch jobs, wait for results	Interactive REPL
Install runtime on every node	Single binary, no deps
Test application performance	Test infrastructure: cooling, power, fabric

stryke is the ultimate load testing tool for distributed computing clusters.

---

[0x11] LANGUAGE SERVER (stryke lsp)

stryke lsp (or stryke --lsp) runs an LSP server over stdio. Hooks into the existing parser, lexer, and symbol table — no separate analyzer to maintain. Surfaces:

• Diagnostics on every keystroke. Strict-vars is on by default in the IDE (CLI stryke check stays lenient — gated on explicit use strict;), so undefined $scalar / @array / %hash typos surface inline without per-file opt-in. Skipped: bare sigil-vars inside double-quoted string interpolations ("got $fh ..." is template/description text — no false positives on test descriptions), Perl $^X/$^O/$^V/$^W family and bare $$ (process id), AOP advice-body context vars ($INTERCEPT_NAME, @INTERCEPT_ARGS, $INTERCEPT_RESULT, $INTERCEPT_MS, $INTERCEPT_US), the full block-param grammar (_, _N, _<, _<N, _<<<<<, _N<<<<<, _N<M), $#arr last-index-of-array references resolved via @arr, open(my $fh, …) lexical filehandle declarations, exists(&Pkg::sub) introspection, and parser-internal _thread_par_run desugaring targets emitted by ~p> / ~s> thread macros.
• OOP-aware checks. Constructor calls (Class(field => v, …) and Class->new(...)) validate keys against declared fields, walking parent classes via extends and implemented traits via impl for inherited fields. Positional constructor calls (Task(1, "title", Priority::High)) are auto-detected and pass through unchanged. $self->method / $obj->method checks walk the same extends + impl chain, plus a universal-method whitelist (isa, can, DOES, does, VERSION, new, BUILD, DESTROY, clone, with, to_hash, to_hash_rec, to_hash_deep, fields, methods, superclass). Match-arm enum-variant typos (Sig::Term2 => when Sig has no Term2) flag with the available-variant list.
• Cross-file require resolution. require "./lib/Foo/Bar.stk" walks up from the source file looking for the project root (any ancestor with a sibling lib/ directory) — the classic Perl/CPAN layout. Subs, classes, traits, enums, constants declared in required files all join the active completion + diagnostics index.
• Hover docs for builtins (pmap, cluster, fetch_json, dataframe, …) — including the parallel and cluster primitives from sections [0x03] and [0x10], every Perl special variable ($!, $@, $_, @_, @ARGV, %ENV, $1..$9, $^A..$^X, @+/@-, …), the __NAME__ compile-time tokens (__END__, __DATA__, __FILE__, __LINE__, __PACKAGE__, __SUB__), every phase block (BEGIN/UNITCHECK/CHECK/INIT/END/BUILD), and the reflection-hash short aliases (%a/%b/%c/%d/%e/%k/%o/%p/%pc/%v, %parameters, %limits, %term). Hash-returning builtins like pool_info() ship full key tables in their hover doc. Hover is suppressed inside string literals ("length" doesn't pop the length builtin) but still fires inside #{ EXPR } interpolations.
• Completion covering every identifier category the editor expects to see:
  • Sigil variables ($scalar, @array, %hash) — declared via my / our / state / local / varsync / oursync, foreach my $x, sub signature params, open(my $fh, …) filehandle decls
  • Subs — bare and qualified, with suffix-only insertText for qualified completions (typing Demo::│ produces Demo::handle, not Demo::Demo::handle)
  • User-declared Types — classes (CompletionItemKind::CLASS), structs (STRUCT), enums (ENUM), traits (INTERFACE), plus every enum variant as a qualified EnumName::Variant
  • Constants from use constant NAME => … and the use constant { A => 1, B => 2 } hash form
  • Loop labels for last LOOP / next LOOP / redo LOOP references
  • Hash-key completion driven by builtin return schemas — val $info = pool_info(); $info->{<tab>} lists the actual keys pool_info returns (cpus, rayon_threads, arch, os, perf_cores, eff_cores). Same registry covers par_bench, stress_test, cache_stats, uname, audio_info, id3_read, git_log, git_show, git_status, git_branches, git_blame, git_authors, du_tree, process_list, net_interfaces, perfview, mounts, html_parse, css_select, xml_parse, xpath. foreach val $row (git_log()) { $row->{<tab>} } also resolves through the loop-var binding.
  • Stryke keywords (fn, class, struct, enum, trait, match, varsync, frozen, …) and ~10k builtins from %all
  • use Module VERSION version slot — use GUI <TAB> / use Foo 1<TAB> lists every installed version directly from ~/.stryke/store/<canonical>@<ver>/. Honors the namespace bridge (use GUI finds versions of stryke-gui) and prefix-filters on what the user typed (1<TAB> shows only 1.x versions). The list comes from the same store-scan the resolver uses, so any offered version is guaranteed to satisfy use Module VERSION at runtime.
  • In-progress parse recovery — when the cursor sits inside a fragment that breaks the parse (Demo::│), the LSP retries with the cursor's line blanked so completion still indexes the rest of the file
  • Trigger characters include { so $h->{ / $h{ auto-popup the relevant key set
• Semantic tokens for server-driven syntax coloring — keywords, builtins (with defaultLibrary modifier), sigil variables, pipe operators, regex literals, numbers, strings, comments — beyond what any client-side lexer can know
• Goto / References / Rename — package-aware. Rename of struct/class/enum/trait fields and methods is AST-based with no textual fallback (so val %h = (width => 1) is not mistaken for a field reference when renaming width). The server defensively strips a :: qualifier from newName so clients that send the full prefilled identifier (e.g. TrafficLight::Stop when renaming a variant) still produce the correct bare replacement. Cross-file rename walks the require graph BFS-style.
• Code actions — Extract Variable / Constant / Parameter / Function (Cmd-Opt-V / Cmd-Opt-C / Cmd-Opt-P / Cmd-Opt-M), Wrap-in-p, toggle line comment. Extract works on caret-only (no manual selection needed) and inside double-quoted strings / backticks.
• Signature help — parameter hints derived from the same doc strings that drive hover; active-parameter tracking as you type past commas

Wire it into VS Code, JetBrains, or any LSP-aware editor by pointing the client at stryke lsp (or stryke --lsp) as the language-server command. There is no separate stryke-lsp binary — the same stryke you run scripts with also acts as its own language server.

// .vscode/settings.json
{
  "stryke.serverPath": "/usr/local/bin/stryke",
  "stryke.serverArgs": ["--lsp"]
}

For JetBrains IDEs (RustRover, IDEA Ultimate, GoLand, PyCharm Pro, WebStorm, RubyMine, PhpStorm, CLion, Rider, DataGrip, Aqua) there is a first-class plugin under editors/intellij/:

• .stk file association, 44-slot color scheme, hand-rolled lexer with finer-grained token categories (declaration vs control keywords, sigil variants, topic / block-param / special-var splits, pipe / arrow / range / regex-bind operators). The lexer correctly handles every Perl regex-operator form as one atomic REGEX token — s/PATTERN/REPL/FLAGS, tr/.../.../, y/.../.../, m/.../, qr/.../ — including embedded quote chars (s/"/""/g), paired-bracket delimiters (s{foo}{bar}), and mixed delimiters. Keyword spellings used as method names (fn state, fn transition) tokenize as FUNCTION_DECL, not the matching control/decl keyword. Keyword spellings used as hash keys ($h->{state}, state => 1) tokenize as IDENTIFIER. Perl-style @{[ EXPR ]} array-ref interpolations inside double-quoted strings tokenize the interior as code. The full block-param grammar (_, _<, _<2, _<<<<<, $_<3, $_2<<<) lexes as single BLOCK_PARAM tokens.
• LSP client wired to stryke --lsp — completion (with 60+ snippets, hash-key completion driven by builtin return schemas, suffix-only insertText for qualified Demo::<tab>, parse-error recovery during typing), hover (full markdown cards, suppressed inside string literals so "length" doesn't pop builtin docs), goto / refs / rename (AST-based for class/struct/enum/trait fields; the rename dialog prefills with the bare segment so editing Red → Redgg doesn't qualify the result as TrafficLight::Redgg), semantic tokens, signature help, code actions (Wrap-in-p, Comment / Uncomment, Extract to variable / constant / parameter / function — caret-only, inside-string aware), folding ranges (every { … } block + =pod ... =cut + 3+ #-line comment runs), document formatting (Cmd-Opt-L pipes through fmt::format_program), diagnostics (strict-vars by default with sensible exemptions — see [0x11])
• Run configurations with --no-interop / --disasm / --profile / --flame / -d toggles; context-menu Run with stryke on any .stk
• Debugger over the Debug Adapter Protocol (stryke --dap): line + function breakpoints from the gutter, Continue / Step Over / Step Into / Step Out / Pause / Run-to-Cursor, frames with file:line, recursive hash & array drill-down in the Variables panel, real-time p / print output in the Console, Evaluate dialog with scalar prelude injection so $a * $b returns the right value from the paused frame. The CLI debugger stryke -d file.stk is a separate TTY front-end on the same Debugger state machine — both share breakpoint / step / scope inspection logic.
• Reflection tool window (View → Tool Windows → Stryke) — searchable trees of %stryke::all, %stryke::builtins, %stryke::keywords, %stryke::operators, %stryke::special_vars, %stryke::perl_compats, %stryke::extensions, %stryke::aliases, %stryke::descriptions (≈25k entries). Left-click a leaf for an ANSI-rendered docs popup; right-click for Show Docs / Copy Name.

Build with ./gradlew buildPlugin and install the zip from editors/intellij/build/distributions/. Community editions are not supported (no LSP API). See editors/intellij/README.md for the full feature list, settings, and architecture.

stryke --dap — Debug Adapter Protocol

The same stryke binary that runs scripts also speaks the standard DAP protocol. Two transport modes:

• Stdio (stryke --dap) — DAP messages over stdin/stdout. Useful for shell testing and clients that drive adapters over their own stdio.
• TCP (stryke --dap HOST:PORT) — connects to a listening socket the client opened first. This is what the JetBrains plugin uses; it keeps stryke's stdout/stderr free for the program's own p/print output (which the IDE's OSProcessHandler reads independently).

Supported DAP requests: initialize, setBreakpoints (line), setFunctionBreakpoints, setExceptionBreakpoints, configurationDone, launch, threads, stackTrace, scopes, variables (recursive — every hash/array gets a variablesReference for drill-down), continue, next, stepIn, stepOut, pause, evaluate, terminate, disconnect. Supported events: initialized, stopped, output, process, thread, exited, terminated.

The TTY CLI debugger (stryke -d) and the DAP server (stryke --dap) share a single Debugger state machine — breakpoints, step modes, scope inspection, call-stack tracking. The only difference is the front-end transport.

---

[0x12] LANGUAGE REFLECTION

stryke exposes its own parser and dispatcher state as plain Perl hashes, so you can enumerate, look up, filter, and pipe over everything the interpreter knows about — no separate API surface to learn, just standard hash ops.

The data is derived at compile time by build.rs from the source of truth: section-commented groups in is_perl5_core / stryke_extension_name (for categories), try_builtin arm names (for aliases), and doc_for_label_text in src/lsp.rs (for descriptions). No hand-maintained list, no stale counts.

Hashes

Eleven hashes; every direct lookup ($h{name}) is O(1). Forward maps:

Long name	Short	Key → Value
%stryke::builtins	%b	primary callable name → category ("parallel", "string", …). Primaries-only — clean unique-op count. No keywords.
%stryke::keywords	%k	stryke language keyword → category ("control", "decl", "exception", "phase", "concurrency", "oo", "operator", "visibility"). Disjoint from %b.
%stryke::operators	%o	symbol operator spelling → category ("arith", "compare", "logical", "bitwise", "assign", "binding", "pipeline", …). Word operators (and/or/eq/cmp) live in %k.
%stryke::special_vars	%v	special variable spelling (sigil included) → category ("error", "regex-capture", "caret", "env", "script", "args", …). One hash covers every kind: $!, @ARGV, %ENV, $^X, __FILE__, etc.
%stryke::all	%all	every name stryke recognizes — %a + %b + %k. Aliases inherit their primary's tag; keywords carry their %k category. Use this for scalar keys %all.
%stryke::perl_compats	%pc	subset of %b: Perl 5 core only, name → category
%stryke::extensions	%e	subset of %b: stryke-only, name → category
%stryke::aliases	%a	alias → canonical primary ($a{tj} → "to_json")
%stryke::descriptions	%d	name → one-line LSP summary (sparse)

Inverted indexes for constant-time reverse queries:

Long name	Short	Key → Value
%stryke::categories	%c	category → arrayref of names ($c{parallel} → [pmap, pgrep, …])
%stryke::primaries	%p	primary → arrayref of its aliases ($p{to_json} → [tj])

Examples

# O(1) direct lookups
stryke 'p $b{pmap}'              # "parallel"
stryke 'p $b{to_json}'           # "serialization"
stryke 'p $pc{map}'              # "array / list"
stryke 'p $e{pmap}'              # "parallel"
stryke 'p $a{tj}'                # "to_json"
stryke 'p $d{pmap}'              # LSP one-liner
stryke 'p $all{tj}'              # "serialization"  (alias resolved via %all)
stryke 'p $k{if}'                # "control"
stryke 'p $k{class}'             # "decl"
stryke 'p $all{while}'           # "control"        (keyword resolved via %all)
stryke 'p scalar @{$c{parallel}}'  # number of parallel ops
stryke '$p{to_json} |> e p'        # every alias of to_json

# total callable spellings (primaries + aliases), one direct count
stryke 'p scalar keys %all'

# see just Perl compats
stryke 'keys %pc |> sort |> p'

# see just stryke extensions
stryke 'keys %e |> sort |> p'

# enumerate a whole category in O(1)
stryke '$c{parallel} |> e p'
stryke '$c{"array / list"} |> e p'

# browse any of them interactively via the pager
stryke 'keys %all |> less'

# frequency table: how many ops per category?
stryke 'val %f; $f{$b{$_}}++ for keys %b; dd \%f'

# find every documented op mentioning "parallel"
stryke 'keys %d |> grep { $d{$_} =~ /parallel/i } |> sort |> p'

# catalog the full reflection surface
stryke 'for val $h (qw(b k all pc e a d c p)) {
         printf "%%%-4s %d\n", $h, scalar keys %$h
       }'

Notes

• Every direct $h{name} lookup is O(1). Filter queries (grep { cond } keys %h) are O(n), but the two inverted indexes (%c, %p) give you O(1) reverse-lookups for the two most common "find names by property" queries.
• Hash sigil namespace is separate from scalars and subs, so %a/%b/%c/%d/%e/%k/%p/%pc don't collide with $a/$b sort specials or the e extension sub.
• Short aliases are value copies of the long %stryke::* names — currently read-only in practice, so the copy never diverges.
• %descriptions is sparse: exists $d{$name} doubles as "is this documented in the LSP?". Undocumented ops still appear in %builtins with a category — they just lack a hover summary.
• A value of "uncategorized" in %builtins means the name is dispatched at runtime but doesn't match any // ── category ── section comment in parser.rs yet — a flag for "add a section header here", not an error.

[0x14] PACKAGE MANAGER

Cargo-shaped manifest + lockfile, hash-pinned, parallel resolver. Single binary surface (stryke ...), no separate cargo-style entry point. Full design in docs/PACKAGE_REGISTRY.md.

# Lifecycle
stryke new myapp                                      # scaffold project at ./myapp/
stryke init                                           # scaffold project in cwd
stryke add http@^1.0 json                             # registry dep
stryke add github.com/OWNER/REPO                      # github-release dep (latest tag, host-triple tarball)
stryke add github.com/OWNER/REPO@v1.0                 # github-release dep pinned to a specific tag
stryke add https://github.com/OWNER/REPO.git          # full URL form also accepted
stryke add ./mylib                                    # path dep (relative)
stryke add ../sibling                                 # path dep (sibling)
stryke add /work/vendored/mylib                       # path dep (absolute)
stryke add mylib --path=../mylib                      # explicit --path override
stryke add http --dev                                 # dev-deps
stryke add criterion --group=bench
stryke remove http                                    # drop dep from manifest
stryke install                                        # resolve + write stryke.lock
stryke install --offline                              # no network; lockfile-only
stryke update [NAME]                                  # re-resolve within manifest constraints, rewrite stryke.lock
stryke upgrade [NAME]                                 # bump stryke.toml pins to latest upstream, then re-resolve
stryke outdated                                       # report deps drifted from their lock pin
stryke audit                                          # check lockfile against advisory feed
stryke tree                                           # print resolved dep graph
stryke info http                                      # show lockfile entry for a dep
stryke vendor                                         # snapshot store deps to ./vendor/
stryke clean [--all]                                  # wipe target/ (and optionally global caches)

# npm-style task runner
stryke run greet                  # execute [scripts] entry "greet"

# Global CLI tools
stryke install -g ../mytool       # link [bin] entries from a path package into ~/.stryke/bin/
stryke uninstall -g mytool
stryke list -g
stryke upgrade -g                 # re-pin all global packages at latest upstream (github releases / path source)

# Registry surface (registry endpoint not deployed yet — stubs return diagnostics)
stryke search NAME
stryke publish [--dry-run]
stryke yank VERSION

Project layout (examples/project/):

myapp/
├── stryke.toml                   # manifest (name, version, deps, [scripts], [bin], [workspace], etc.)
├── stryke.lock                   # exact versions + integrity hashes (commit this)
├── main.stk                      # entry point (`stryke main.stk` or just `stryke`)
├── lib/                          # module sources, accessed via require/use
├── bin/                          # additional executables (auto-discovered)
├── t/                            # tests (`stryke test t/`)
├── benches/                      # benchmarks (`stryke bench`)
└── target/                       # build outputs (gitignored)

Workspaces are first-class:

# stryke.toml at workspace root
[workspace]
members = ["crates/*"]

[workspace.deps]
shared = { path = "../shared" }   # one version pinned for the whole monorepo

Then in any member's stryke.toml:

[deps]
shared = { workspace = true }     # inherit version + features from the root

Single stryke.lock at the workspace root pins every member's transitive graph.

Deps live globally in ~/.stryke/store/name@version/ — no node_modules/-shaped per-project dependency tree. Every dep is hash-pinned in the lockfile (Nix-style reproducibility, Cargo-style ergonomics). stryke build --release AOT-compiles the whole program — your code, every dep, the stdlib — through Cranelift to a single statically-linked native binary. SFTP-able. No interpreter needed on the target machine.

Status: path deps + workspace deps + GitHub-release deps ({ github = "OWNER/REPO" [, version = "..."] }, downloads the prebuilt host-triple tarball with SHA-256 verification — this is the path FFI cdylib packages like the 14 stryke-* connectors take) + source-clone git deps ({ git = "...", branch|tag|rev = ... }, for pure-source dependencies hosted anywhere) + full CLI surface (new/init/add/remove/install/update/outdated/audit/tree/info/vendor/clean/run/install -g/upgrade -g etc.) are wired and tested today. s add github.com/OWNER/REPO[@TAG] auto-writes the github = "..." form; s add ./PATH auto-writes the path = "..." form. Registry deps + the PubGrub semver resolver land when the registry endpoint is deployed — the CLI stubs for search/publish/yank already exist and return clear "registry not deployed yet" diagnostics so the surface matches the RFC end-state.

Skipped on purpose: install-time code execution (no build.rs / postinstall), hoisting, phantom deps, peer deps, mutable registries. The lockfile is sacred; regenerate explicitly.

---

[0x15] WEB FRAMEWORK (s_web)

Rails-shaped framework that lives in the sibling crate stryke_web/. Generator emits .stk source files; framework runtime is web_* builtins in the main strykelang/ crate. Full reference in stryke_web/README.md.

One-line full-stack app:

s_web new mega --app everything --theme cyberpunk --auth --admin --docker --ci --pwa --migrate
cd mega && bin/server
# ~70 resources, ~490 CRUD routes, dark cyberpunk CSS, signup/login/sessions,
# admin panel at /admin, /health endpoint, Dockerfile, GitHub Actions CI, PWA
# manifest + service worker. Runs at http://localhost:3000.

Component	What's wired
Routing	web_route VERB " /path", "ctrl#act", web_resources "posts" (7-route REST), web_root "ctrl#act", OpenAPI 3.0 dump auto-served at /openapi.json, Swagger UI at /docs
Controllers	web_render(html|text|json|template|redirect), web_params, web_request, web_set_header, web_status, web_security_headers, default-convention render
Views	ERB engine (<%= %> / <% %> / <%# %> / <%- -%>), layout wrap, partials (web_render_partial), web_link_to, web_form_with, web_text_field/area/check_box, web_csrf_meta_tag
ORM (SQLite)	class Article extends ApplicationRecord with auto-generated Self.all/find/where/create/update/destroy static methods. web_model_paginate/search/soft_destroy/count/first/last/with for n+1 elimination, soft delete, pagination
Migrations	web_create_table/drop_table/add_column/remove_column schema DSL, web_migrate/rollback runner, schema_migrations tracking
Validations / strong params	web_validate(+{title => "presence,length:1..100", email => "format:^.+@.+$"}), web_permit($params, "title", "body")
Auth	web_password_hash/verify, web_session_set/get, signed time-limited tokens (web_token_for/consume), CSRF meta, web_can("posts.edit", $user) permissions
Filters	web_before_action/web_after_action with only/except
HTTP cache	web_etag with auto-304 short-circuit, prompt-cache headers
Helpers	web_h (HTML escape), web_truncate, web_pluralize, web_time_ago_in_words, web_image_tag, web_button_to
API	--api mode, s_web g api Post JSON controllers, JSON:API helpers (web_jsonapi_resource/collection/error), web_bearer_token
Themes	9 baked-in: simple, dark, pico, bootstrap, tailwind, cyberpunk, synthwave, terminal, matrix
DevOps	--docker (multi-stage Dockerfile + .dockerignore), --ci (GitHub Actions), --pwa (manifest.json + service worker)
Fat binary	s_web build --out dist && cd dist && cargo build --release produces a single self-contained binary that include_str!s every .stk file plus the stryke runtime — drop on any Linux box, run, no deps
Generators	s_web g {scaffold, model, migration, controller, app, auth, admin, api, mailer, job, channel, docker, ci, pwa}

Presets: blog (8 resources), ecommerce (15), saas (12), social (10), cms (12), forum (10), crm (10), helpdesk (8), plus named clones: amazon (25), facebook (23), learning (21 — Anki-style with SRS), and everything (~70 resources unioned + dedup'd).

---

[0x16] AI PRIMITIVES

ai is a builtin like print — two letters, ubiquitous, unlimited power. Full design + phase-by-phase status in docs/AI_PRIMITIVES.md.

val $r = ai "summarize this", $document       # bare call
val $r = ai "research X", tools => [...]      # auto-routes to agent loop
val $r = ai "describe", image => "/img.jpg"   # vision
val $r = ai "extract", schema => +{...}       # structured output
val $r = ai "...", pdf => "/contract.pdf"     # document input
for val $chunk in stream_prompt("write a haiku") { print $chunk }   # iter-context streaming

Surface	Builtins
Single-shot	ai, prompt, stream_prompt, chat, embed, tokens_of, ai_estimate
Agent loop	ai($p, tools => [...]) — Anthropic tool_use + OpenAI function-calling protocols. Multi-turn, multi-tool, max_turns/max_cost ceilings
tool fn keyword	tool fn weather($city: string) "Get weather" { ... } — auto-schemas signature, auto-registers, auto-attaches to bare ai($p) calls
Built-in tools	web_search_tool (Brave/DDG), fetch_url_tool, read_file_tool, run_code_tool (sandboxed Python) — drop into tools => [...]
MCP client	mcp_connect("stdio:CMD") and mcp_connect("https://..."), full tools/resources/prompts/call surface, auto-attach to agent loop via mcp_attach_to_ai
MCP server	mcp_server_start("name", +{tools => [...]}) runtime, plus declarative mcp_server "name" { tool foo($a) "..." {...} } parser DSL
Sessions	ai_session_new/send/history/reset/close — multi-turn chat tracking
Collection ops	ai_filter, ai_map, ai_classify, ai_match, ai_sort, ai_dedupe — single batched LLM call across the collection
Memory / RAG	ai_memory_save/recall/forget/count/clear — sqlite-backed embedding store, cosine retrieval
Vector ops	vec_cosine, vec_search, vec_topk
Multimodal	ai_vision (image), ai_pdf (document)
Cost / cache	ai_cost, ai_cache_get/set/clear, ai_history, ai_budget($usd, fn { ... }) scoped cap
Mock / test	ai_mock_install, STRYKE_AI_MODE=mock-only for CI
Convenience	ai_summarize, ai_translate, ai_extract, ai_template, ai_last_thinking
Audio	ai_transcribe "audio.mp3" (Whisper), ai_speak "text", voice => "alloy" (OpenAI TTS)
Image	ai_image $prompt, ai_image_edit $prompt, image => $src, mask => $m, ai_image_variation image => $src, n => 4 — DALL-E 3 / gpt-image-1 / DALL-E 2
Catalog	ai_models("openai"|"anthropic"|"ollama"|"gemini") — live model IDs from each provider's /models endpoint
Citations	ai_pdf $p, pdf => $f, citations => 1 and ai_grounded $p, documents => [@paths] — multi-doc grounding with auto-citations via ai_citations()
Files (OpenAI)	ai_file_upload "file.bin", purpose => "user_data", ai_file_list, ai_file_get, ai_file_delete
Files (Anthropic)	ai_file_anthropic_upload "file.pdf", ai_file_anthropic_list, ai_file_anthropic_delete — beta files-api-2025-04-14
Moderation	ai_moderate $text → +{ flagged, categories, scores } — OpenAI safety classifier (free endpoint)
Chunk	ai_chunk $text, max_tokens => 500, overlap => 50, by => "chars"|"sentences" — RAG primitive, no API call
Warm / verify	ai_warm(model => ..., provider => ...) → +{ ok, latency_ms, error } — auth + reachability ping at script start
Compare	ai_compare $a, $b, criteria => "...", scale => 5 → +{ winner, reason, scores } — single-call structured comparison
Dashboard	print ai_dashboard() — ANSI summary of cost/tokens/cache hit-ratio
Pricing	ai_pricing("claude-opus-4-7") → +{ input, output, input_per_1m, output_per_1m } for pre-flight cost estimates
Describe	ai_describe "img.png", style => "concise"|"detailed"|"alt" — vision wrapper with style presets
Sessions	ai_session_new/send/history/reset/close plus ai_session_export($h) → $json and ai_session_import($json) → $h for persistence across runs
Embed providers	Voyage (default), OpenAI (text-embedding-3), Ollama (nomic-embed-text/mxbai-embed-large — local, $0/M tokens)
CLI modes	stryke ai --image PROMPT -o out.png, --transcribe audio.mp3 -o out.txt, --speak "hello" -o out.mp3 — UNIX-filter mode covers chat, image, audio
Batch	ai_batch(\@prompts) — Anthropic batch API at 50% cost
Cluster fanout	ai_pmap(\@items, "instruction", cluster => $c) — distributed AI work
CLI	stryke ai "prompt" — UNIX filter mode with --model, --system, --stream, --json

Providers wired: Anthropic (full surface incl. extended thinking, prompt caching, vision, PDF, batch), OpenAI (Chat + tool calls + streaming, Whisper, TTS), Voyage (embeddings, default), Ollama (/api/generate), OpenAI-compatible (openai_compat/compat/local — LM Studio, vLLM, llama-server; configurable STRYKE_AI_BASE_URL), Google Gemini. In-process llama.cpp deferred — Ollama / LM Studio is the supported local path today.

# Auto-attached: bare `ai()` sees the tool fn without `tools =>` arg.
tool fn current_user($username: string) "Look up a user" {
    User::find_by_email($username)
}

tool fn create_post($title: string, $body: string) "Create a post" {
    Post::create(+{ title => $title, body => $body })
}

val $reply = ai("create a post titled 'Hello' from alice@x.io with body 'World'");

---

[0x17] EXPECT / INTERACTIVE AUTOMATION

PTY-driven interactive scripting — the modern Tcl/Expect successor. Full design + phase status in docs/expect-feature-idea.md.

val $h = pty_spawn("ssh user@host");
pty_expect($h, qr/password:/, 30);
pty_send($h, "hunter2\n");
pty_expect($h, qr/\$ /, 30);
pty_send($h, "uptime\n");
val $out = pty_expect($h, qr/\$ /, 30);
pty_close($h);

# Table form (Tcl `expect { ... }` block, in stryke):
val $tag = pty_expect_table($h, [
    +{ re => qr/password:/, do => fn { pty_send($h, "$pw\n"); "ok" } },
    +{ re => qr/yes\/no/,   do => fn { pty_send($h, "yes\n"); "confirmed" } },
    +{ re => qr/denied/,    do => fn { die "auth failed" } },
], 30);

# Method-form sugar (require "perl_pty_class.stk"):
val $h = PtyHandle::spawn("ssh host");
$h->expect(qr/password:/, 30);
$h->send("$pw\n");
$h->branch([+{re => qr/\$ /, do => fn { "shell ready" }}], 30);
$h->interact();   # raw-mode handoff, Ctrl-] to detach
$h->close();

Builtin	Behavior
pty_spawn(cmd) / pty_spawn(cmd, arg, ...)	Allocate PTY via nix::pty::openpty, fork+exec child, return handle
pty_send($h, "text")	Write to master fd
pty_read($h, timeout_secs)	One-shot read, returns string or undef on EOF
pty_expect($h, qr/.../, timeout?)	Loop: try regex on buffer, else select() + drain, retry until match or timeout
pty_expect_table($h, [+{re, do}, ...], timeout?)	Multi-pattern dispatch — first match wins; calls the matched branch's do sub
pty_buffer($h) / pty_alive($h) / pty_eof($h)	Inspection
pty_close($h)	SIGTERM → 200ms grace → SIGKILL, returns exit status
pty_interact($h)	Raw-mode handoff: tcgetattr/cfmakeraw, select() on stdin+master, forward both directions until EOF or Ctrl-]

Combined with pmap_on cluster dispatch you get parallel SSH automation across N hosts:

val $cluster = cluster(["host1:8", "host2:8", "host3:8"]);
pmap_on $cluster @hosts -> $host {
    val $h = pty_spawn("ssh $host");
    pty_expect($h, qr/password:/, 10);
    pty_send($h, "$passwords{$host}\n");
    pty_expect($h, qr/\$ /, 30);
    pty_send($h, "apt update && apt upgrade -y\n");
    pty_expect($h, qr/\$ /, 1800);
    pty_close($h);
}

Unix-only for v0; Windows ConPTY support is a separate code path that's still pending.

---

[0x18] DOCUMENTATION

All documentation is served via GitHub Pages at menketechnologies.github.io/strykelang/.

Document	Description
Docs Home	Stryke reference — quickstart, builtins, parallel primitives, pipe-forward syntax, reflection hashes
Full Reference	Complete language reference — every builtin, operator, special variable, and regex feature
Engineering Report	strykelang internals — Rust lines, callable builtins, VM opcodes, AST variants, Cranelift JIT, rayon-backed parallel runtime, perl-parity cases
PACKAGE_REGISTRY.md	Package manager design — manifest, lockfile, resolver, global store, build outputs
stryke_web/README.md	Web framework — generators, presets, themes, builtins, fat-binary build
AI_PRIMITIVES.md	AI primitives — agent loop, MCP, tool fn, RAG, vector search, providers, phase-by-phase status
expect-feature-idea.md	Interactive automation — PTY runtime, expect tables, cluster fanout
STRESS_TESTING.md	Distributed load testing — stress_* builtins, agent/controller, hardware/cooling validation
WEB_FRAMEWORK.md	Original web-framework design RFC
ROADMAP.md	Forward-looking work and open questions

---

[0xFF] LICENSE

MIT — see LICENSE.

---

░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
░░ >>> PARSE. EXECUTE. PARALLELIZE. OWN YOUR CORES. <<< ░░
░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░

created by MenkeTechnologies