Previous version ran all 7 claude sessions in the main working tree on
branch 'architecture'. That would race on git operations and cross-
contaminate commits between languages even though their file scopes
don't overlap. Now each session runs in /root/rose-ash-loops/<lang> on
branch loops/<lang>, created from the current architecture HEAD.
sx-loops-down.sh gains --clean to remove the worktrees; loops/<lang>
branches stay unless explicitly deleted.
Also: second Enter keystroke after the /loop command, since Claude's
input box sometimes interprets the first newline as a soft break.
sx-loops-up.sh spawns a tmux session 'sx-loops' with 7 windows (lua,
prolog, forth, erlang, haskell, js, hs). Each window runs 'claude'
and then /loop against its briefing at plans/agent-briefings/<x>-loop.md.
Optional arg is the interval (e.g. 15m); omit for model-self-paced.
Each loop does ONE iteration per fire: pick the first unchecked [ ] item,
implement, test, commit, tick, log — then stop. Commits push to
origin/loops/<lang> (safe; not main).
sx-loops-down.sh sends /exit to each window and kills the session.
Attach with: tmux a -t sx-loops
- scripts/loop-guard.sh — atomic claim with 30-min staleness overtake,
appends NDJSON event to .loop-logs/<lang>.ndjson. Exit 0 = go ahead,
exit 1 = another run is live, skip.
- scripts/loop-release.sh — clear lock, log release with exit status.
Intended for 7 per-language /schedule routines firing every 15 minutes.
Lock detects overlap so tight cadences are safe; stale lock (>30 min)
overtaken automatically if an agent dies mid-run.
sx-tree MCP file ops broken this session (Yojson Type_error "Expected
string, got null" on every file-based call — sx_read_subtree,
sx_find_all, sx_replace_by_pattern, sx_summarise, sx_pretty_print, even
sx_load_check on existing files works but summarise fails). Can't edit
integration.sx to add before:init/after:init dispatch. Additionally 4
of the 6 tests fundamentally require stricter parser error-rejection
(add - to currently parses to (set! nil ...); on click blargh end
accepts blargh as symbol expression) — out of single-cluster budget.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cluster 19 was implemented in 4be90bf2 but the plan/scoreboard rows
still marked it pending. Sync the plan state: mark done, add log entry,
bump merged total 1264 → 1277.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Five new guest-language plans mirroring the js-on-sx / hs-loop pattern, each
with a phased roadmap (Progress log + Blockers), a self-contained agent
briefing for respawning a long-lived loop, and a shared restore-all.sh that
snapshots state across all seven language loops.
Briefings bake in the lessons from today's stall debugging: never call
sx_build (600s watchdog), only touch lib/<lang>/** + own plan file, commit
every feature, update Progress log on each commit, route shared-file
issues to Blockers rather than fixing them.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds regex-compile/test/exec/match-all/replace/replace-fn/split/source/flags.
Opaque dict handle {:__regex__ true :id :source :flags}; compiled Re.re
cached in a primitives-local table. Replacement supports $&, $1-$9, $$.
Flags: i (CASELESS), m (MULTILINE), s (DOTALL). g is a runtime flag handled
in replace. u (unicode) skipped for now.
Unblocks js-on-sx's regex-platform-override! hook — the JS RegExp shim can
now delegate to real regex instead of the substring stub.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Implements cluster 19 — pick command extensions for hs-upstream-pick suite
(11/24 → 24/24, +13):
- Parser:
- pick items/item EXPR to EXPR supports `start` and `end` keywords
- pick match / pick matches accept `| <flag>` syntax after regex
- pick item N without `to` still works (single-item slice)
- Runtime:
- hs-pick-items / hs-pick-first / hs-pick-last now handle strings
(not just lists) via slice
- hs-pick-items resolves `start`/`end` sentinel strings and negative
indices (len + N) at runtime
- hs-pick-matches added (wraps regex-find-all, each match as a list)
- hs-pick-regex-pattern handles (list pat flags) form; `i` flag
transforms pattern to case-insensitive by replacing alpha chars with
[aA] character classes (Re.Pcre has no inline-flag support)
- Generator:
- extract_hs_expr now decodes JS string escape sequences (\" -> ",
\\ -> \) instead of stripping all backslashes, then re-escapes for
SX. Preserves regex escapes (\d, \s), CSS escapes, and lambda `\`
syntax for String.raw template literals while still producing
correct output for regular JS strings.
Smoke (0-195): 170/195 unchanged (no regressions).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Shared formatter in sx_types.ml. Small integer-valued floats still print
as plain ints; floats outside safe-int range (|n| >= 1e16) now print as
%.17g (full precision) instead of silently wrapping to negative or 0.
Non-integer values keep %g 6-digit behavior — no existing SX tests regress.
Unblocks Number.MAX_VALUE / Math.pow(2,N) style tests in js-on-sx where
iterative float loops were collapsing to 0 at ~2^63.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Number.prototype.constructor === Number etc. Four dict-set! lines add
the backlink after each constructor dict is defined.
new String().constructor === String now returns true. Array literals
don't yet link to Array.prototype so [].constructor === Array is still
false — that would need a boxing refactor.
Unit 521/522, slice 148/148 unchanged.
Number 76/100 → 77/100 (+1). String variance-heavy under CPU load.
Wire up the `ask` and `answer` commands end-to-end:
- tokenizer.sx: register `ask` and `answer` as hs-keywords.
- parser.sx: cmd-kw? gains both; parse-cmd dispatches to new
parse-ask-cmd (emits `(ask MSG)`) and parse-answer-cmd, which
reads `answer MSG [with YES or NO]`. The with/or pair reads
yes/no via parse-atom — parse-expr would collapse
`"Yes" or "No"` into `(or "Yes" "No")` before match-kw "or"
could fire. The no-`with` form emits `(answer-alert MSG)`.
- compiler.sx: three new cond branches (ask, answer, answer-alert)
compile to a let that binds __hs-a, sets `the-result` and `it`,
and returns the value — so `then put it into ...` works.
- runtime.sx: hs-ask / hs-answer / hs-answer-alert call
window.prompt / confirm / alert via host-call + host-global.
- tests/hs-run-filtered.js: test-name-keyed globalThis.{alert,
confirm,prompt}; __currentHsTestName is updated before each
test. Host-set! for innerHTML/textContent now coerces JS
null → "null" (browser behaviour) so `prompt → null` →
`put it into #out` renders literal text "null", which the
fourth test depends on.
Suite hs-upstream-askAnswer: 1/5 -> 5/5.
Smoke 0-195: 166/195 -> 170/195.
Three parser additions so scripts like `(on click (log me) (trigger foo))`
parse into a single feature with both commands in its body:
1. parse-feat: new cond branch for `paren-open` — advance, recurse
parse-feat, consume `paren-close`. Allows a feature like `(on click
...)` to be grouped in parens.
2. parse-cmd: two new cond branches — on `paren-close` return nil (so
cl-collect terminates at an outer group close), and on `paren-open`
advance / recurse / close. Allows single parenthesized commands like
`(log me)`.
3. cl-collect: previously only recursed when the next token was a
recognised command keyword (`cmd-kw?`), so after `(log me)` the
sibling `(trigger foo)` would end the feature body and re-surface as
a top-level feature. Extended the recursion predicate to also fire
when the next token is `paren-open`.
Suite hs-upstream-core/parser: 9/14 -> 10/14.
Smoke 0-195: 165/195 -> 166/195.
SX's (str 1e-7) gives "1e-07" but JS spec is "1e-7" — no padding, no
leading zeros in the exponent (sign stays). We stepped through:
mant "e" expraw → mant "e" (sign (strip-zeros body))
Added four small helpers: js-normalize-num-str, js-split-sign,
js-strip-leading-zeros, js-strip-zeros-loop. All pure string walkers.
Unit 521/522, slice 148/148 unchanged.
String 40 → 42, Number 75 → 76 (+3 total).
Fixes S9.8.1_A9_T1, fromCharCode/S9.7_A3.1_T1..T2 family.
Tests expected Function.prototype.isPrototypeOf(Number/String/…) ===
true because every built-in ctor inherits from Function.prototype.
Our model doesn't link Number.__proto__ anywhere, so the default
Object.isPrototypeOf walked an empty chain and returned false.
Fix: post-definition dict-set! adds an explicit isPrototypeOf override
on js-function-global.prototype that returns (js-function? x) — which
accepts lambdas, functions, components, and __callable__ dicts. Good
enough to satisfy the spec for every case that isn't a bespoke proto
chain.
Unit 521/522, slice 148/148 unchanged.
Wide scoreboard: 156/300 → 159/300 (+3, Number/S15.7.3_A7 and the
three S15.5.3_A2 / S15.6.3_A2 / S15.9.3_A2 twins).
Every built-in JS function on Math/Number/Array/Object had .name === ""
because js-invoke-function-method/js-get-prop returned bare "" for the
"name" slot. That breaks tests like Math.abs.name === "abs" and
Array.isArray.name === "isArray".
Fix: extract the SX symbol name from (inspect fn) which prints
<js-math-abs(x)>, then unmap through a small string table that maps
js-math-abs → "abs", js-array-is-array → "isArray" etc. Also strips
the angle-bracket marker and stops at ( or space.
Non-mapped lambdas (user fns) fall through to the raw "js-foo" form
rather than "", which is slightly worse but only hit in debug prints.
Unit 521/522, slice 148/148 unchanged.
Scoreboard: Math 40/100 → 43/100 (+3); Number 74 → 75 (+1).
Sample: Math/abs/name.js, Math/floor/name.js, Math/max/name.js,
Number/isNaN/name.js — all flipped. length.js tests still fail for
trig because the underlying fn isn't implemented.
js-number-to-string did (str n), which gives OCaml-native "inf"/"-inf"/
"nan"/"-nan" strings. JS spec requires "Infinity"/"-Infinity"/"NaN".
Fix: cond-check js-number-is-nan, and =infinity-value first, fall
through to (str n) for finite.
Unit 521/522, slice 148/148 unchanged.
String scoreboard: 34/100 → 38/100 (+4, S15.5.1.1_A1_T11/T12 family —
String(1/0)/String(-1/0)/String(0/0)).
Previously a JS function body with no return fell through the call/cc
begin as nil, making `(function(){}())` return null and typeof → object.
Spec: falls-off-end gives undefined.
Wrap the call/cc in (let ((__r__ ...)) (if (= __r__ nil) :js-undefined __r__)).
Downside: explicit `return null` also returns nil, but so does (pick
your last expression evaluating to null). For 99% of cases it's
fall-off-end and the fix is correct. Code that genuinely needs
distinguishable null would need separate nil/undef handling in the
evaluator.
Unit 521/522, slice 148/148 unchanged.
Number 73/100 → 74/100 (+1), String 33/100 → 34/100 (+1).
Fixes S15.5.1.1_A1_T1 family (String(function(){}()) should be "undefined").
ES spec: ToNumber("0x0") == 0, ToNumber("0xFF") == 255. Our parser
returned NaN for anything with a 0x/0X prefix, failing S9.3.1_A16..A32
(every hex-literal assertion test case).
Added:
- js-hex-prefix? — is this an 0x/0X prefix?
- js-is-hex-body? — all remaining chars are [0-9a-fA-F]?
- js-parse-hex — walk chars, accumulate in base 16, NaN on bad char
- js-hex-digit-value — char → 0..15 (or -1)
js-is-numeric-string? short-circuits to the hex-body check on 0x*
prefix; js-num-from-string dispatches to js-parse-hex on same.
Unit 521/522, slice 148/148 unchanged.
Number scoreboard: 58/100 → 73/100 (+15).
Sample flipped: S9.3.1_A16 (0x0/0X0), A17..A31 (0x0..0xF and 0x10..0x1F
range coverage). Many of the remaining 22 fails are (new Number()).x
style prototype-chain introspection and MAX_VALUE precision.
js-get-ctor-proto used to always synthesise a per-ctor-id empty dict in
__js_proto_table__, ignoring the :prototype slot that every built-in
constructor dict (Number, String, Array, Boolean, Object, Function)
already carries. So `new Number()` got `{__proto__: {}}` instead of
`{__proto__: Number.prototype}`, breaking every prototype-chain-method
lookup:
(new Number()).toLocaleString // undefined → function
(new Number()).toLocaleString === Number.prototype.toLocaleString
Fix: when the receiver is a dict with a "prototype" key, return that
directly; otherwise fall through to the existing proto-table path.
Declared classes from JS still go through the table because they emit
js-set-ctor-proto! at definition.
Unit 521/522, slice 148/148 unchanged.
Number scoreboard: 53/100 → 58/100 (+5). Sample: S15.7.5_A1_T03..T07
(toLocaleString/toString/toFixed/toExponential prototype identity).
Six post-definition dict-set! calls add .length=1 and .name="<Ctor>"
to the global constructor dicts. Per spec Number/String/Array/Boolean/
Object all have length 1 (they take one arg).
Unit 521/522, slice 148/148 unchanged.
Number scoreboard: 52/100 → 53/100 (+1, S15.7.3_A8: Number.length==1).
Root cause of many Number/String/Object.hasOwnProperty false-negatives:
global dicts like Number/String/Object carry a :__callable__ slot so
they can be invoked (Number(5) coerces, Array(3) makes length-3 list).
That makes js-function? return true for them, so js-invoke-method
dispatched hasOwnProperty/isPrototypeOf/propertyIsEnumerable through
js-invoke-function-objproto (whose name/length/prototype-only check
returns false for real dict keys like MAX_VALUE).
Fix: in the invoke-method cond, exclude dicts from the function-proto
branch. Callable dicts fall through to js-invoke-object-method, which
walks (keys recv) properly.
One line in a compound `and` — minimal surface, easy to revert.
Unit: 521/522 unchanged.
Conformance: 148/148 unchanged.
Number scoreboard: 46/100 → 52/100 (+6).
Impacted sample: Number.hasOwnProperty("MAX_VALUE") → true (was false),
plus S15.7.3_A2..A8 family (MAX_VALUE/MIN_VALUE/POSITIVE_INFINITY/
NEGATIVE_INFINITY existence checks) and S15.7.2.1_A2..A4.
js-num-from-string now finds an e/E split, parses mantissa and exponent
separately, and combines via js-pow-int (positive-exp loop for >=0, 1/
reciprocal for negative). Previously `.12345e-3` parsed as 0.12345 and
"1e3" returned NaN — the parser walked decimals/dots only.
New helpers:
- js-find-exp-char / -loop : linear scan for e/E, returns -1 if absent
- js-pow-int base exp : integer-exp power, handles negative
Also fixed `js-string-trim` typo → `js-trim` in the rewritten num-from-
string, and corrected test 903's expected part count (3, not 2 — the
lexer has always split `hi ${x}!` into str+expr+str, the test just had
the wrong count).
Unit: 521/522 (was 520/522, 934 still blocked on SX \` escape).
Conformance: 148/148 unchanged.
Number scoreboard: 43/100 → 46/100 (+3).
Impacted test262 paths (sample): built-ins/Number/S9.3.1_A11.js and
A12/A16/A17 (".12345e-3", scientific notation round-trips).
Two shared-file entries based on scoreboard patterns:
- Math trig/transcendental primitives missing. 34× "TypeError: not a
function" across Math category — sin/cos/tan/asin/acos/atan/atan2,
sinh/cosh/tanh/asinh/acosh/atanh, log/log2/log10/log1p/expm1,
clz32/imul/fround, variadic hypot/max/min. All need OCaml/JS platform
primitives; can't polyfill from pure SX and keep precision. Once
present in the runtime, `js-global.Math` gets one extension and all 34
failures flip together.
- Evaluator CPU bound at ~1 test/s on 2-core box. Runner already
auto-disables parallel workers on ≤2 cores. Optimization surface for
the shared evaluator: lexical addresses (vs name walk), inline caches
on js-get-prop (vs __proto__ walk), force-JIT transpiled JS bodies
(vs lazy), OCaml 5 domains (vs separate processes).
Progress-log entry for P0 harness cache added alongside.
Root cause: every sx_server worker session used js-eval on the 3.6KB
HARNESS_STUB, paying ~15s for tokenize+parse+transpile even though every
session does the same thing. Over a full scoreboard with periodic worker
restarts that's minutes of wasted work.
Fix: transpile once per Python process. Spin up a throwaway sx_server,
run (inspect (js-transpile (js-parse (js-tokenize HARNESS_STUB)))), write
the resulting SX source to lib/js/.harness-cache/stub.<fingerprint>.sx and
a stable-name symlink-ish copy stub.sx. Every worker session then does a
single (load .harness-cache/stub.sx) instead of re-running js-eval.
Fingerprint: sha256(HARNESS_STUB + lexer.sx + parser.sx + transpile.sx).
Transpiler edits invalidate the cache automatically. Runs back-to-back
reuse the cache — only the first run after a transpiler change pays the
~15s precompute.
Transpile had to gain a $-to-_js_dollar_ name-mangler: the SX tokenizer
rejects $ in identifiers, which broke round-tripping via inspect. JS
$DONOTEVALUATE → SX _js_dollar_DONOTEVALUATE. Internal JS-on-SX names are
unaffected (none contain $).
Measured: 300-test wide (Math+Number+String @ 100/cat, --per-test-timeout 5):
593.7s → 288.0s, 2.06x speedup. Scoreboard 114→115/300 (38.3%, noise band).
Math 40%, Number 44%, String 30% — same shape as prior.
Baselines: 520/522 unit, 148/148 slice — unchanged.
