Anthropic now reserves /plan as a UI command, so the Skill tool refuses to
dispatch it. Rename the project's plan skill to planner, update the symlink
under .claude/skills/, and adjust references in CLAUDE.md, DESIGN.md,
skills/README.md, and the cross-references between brainstorm / implement /
audit / fieldtest / fieldtester. Plan files themselves (docs/plans/*.md)
keep their name — only the skill ID changes.
Reversal of the language decision from 6d94fa5. Texts that get
committed to the repo permanently should follow the same English-
only rule as everything else; the Notify push stays 1:1 with the
WhatsNew entry, so both are English now.
The "only file intentionally not in English" carve-out in CLAUDE.md
is removed.
Done-state notifications now produce two synchronised outputs sharing
one text: an entry appended to docs/WhatsNew.md and the same string
sent via notify.sh. The text is written in German for the user-as-
reader who did not watch the session — no technical internals, no
iteration codes, lead with the change-in-the-project.
WhatsNew.md is the only file in the repo intentionally not in English;
the exception is documented in CLAUDE.md (Roles of docs/...) so a
future reader does not "correct" it back. Bounce-back notifications
stay Notify-only.
Adds a priority-ordered, checkbox-format roadmap that the
orchestrator maintains. Three states (`[ ]` open / `[~]` in
progress / `[x]` done), four kinds (milestone / feature / todo /
idea), four priorities (P0..P3). Initial population lifts the
queued items from the JOURNAL tail.
JOURNAL's role narrows: chronological decisions log only, no
forward queue. CLAUDE.md "Roles of ..." section updated to
reflect the split.
Replace the original "Invent your own programming language" prompt
(the very first text in the repo) with a focused statement of the
language's design goal: AILang is for LLM authors, contrary to
conventional compiler design. Lead with the four design priorities
(machine readability, local reasoning, provability, hallucination
robustness) and the cut/keep asymmetry from DESIGN.md's
feature-acceptance criterion.
Compress three sections that now duplicate skill content:
- Skill-system bullet list -> 1-line pointer to skills/README.md.
- Bug-fixes TDD headline -> 1-line pointer to skills/debug/, as
required by skills/debug/SKILL.md ("CLAUDE.md keeps a one-line
pointer").
- Milestone-cycle pipeline diagram + tidy mandate -> short pointer
to skills/README.md and skills/audit/SKILL.md, retaining only the
iter/family vocabulary note that lives nowhere else.
Also fix the Code-layout skills row to match the current repo
(fieldtest had been missing).
Discipline sections that govern behaviour *between* skills (orchestrator
role, when-not-to-delegate, design-rationale-vs-effort, direction
freedom, notifications) and substantive always-on rules (LLM-utility
feature gate) stay verbatim — they are not skill-replaceable.
247 -> 229 lines.
The four-phase debug process and the five-phase fieldtest process
each lived twice — once in SKILL.md (orchestrator-facing) and once
in the dispatched agent's file (subagent-facing). The orchestrator
does not execute these phases; the subagent does. Duplicating them
in SKILL.md just bloated the orchestrator's main context with bytes
that only the subagent ever needs.
SKILL.md now carries only what the orchestrator must consult at
dispatch time: trigger gating, Iron Law as headline, the carrier
contract, the produced handoff, and cross-references. Iron Law in
operational form, full process, Common Rationalisations, Red Flags,
and (for fieldtest) the spec template now live solely in the agent
file. Net −160 LOC across the skills tree.
CLAUDE.md pointer for 'Bug fixes — TDD, always' updated to reflect
that the substantive discipline lives in the agent file, not the
skill file.
agents/ contained only README.md after the 2026-05-09 migration. The
roster information is more naturally located alongside the skills it
points into. skills/README.md is now the single index (skill table,
pipeline diagram, agent roster, discovery, conventions, how-to-add).
CLAUDE.md updated: code-layout entry consolidated, @-reference points
to skills/README.md.
Adds DESIGN.md "Feature-acceptance criterion" as a top-level section:
a feature ships only if (1) an LLM author naturally produces code
that uses it, and (2) it measurably improves correctness or removes
redundancy. Aesthetic appeal and human ergonomics do not count.
Mirrored in CLAUDE.md as "Feature acceptance: LLM utility", paired
with the existing "Design rationale != implementation effort". The
two together narrow valid feature rationales to one thing: what the
LLM author gets out of the feature.
Trigger: the typeclass-design conversation around 22a. Rule was
implicit in many past decisions (Decision 1's JSON-over-text choice,
Decision 10's "what LLMs are good at" reasoning) but never stated
as a feature-gate. Codifying it now means future feature proposals
get evaluated against an articulated criterion instead of being
re-derived each time.
Documentation-only; no Rust, schema, or bench changes. Test state
288/0/3 unchanged.
Closes the question CLAUDE.md has carried since day one ("LLVM-
linkable, performance is extremely important") with data. Hand-C
variants of the four bench fixtures, compiled with clang -O2,
each carefully matching the AILang algorithm and explicitly
documenting representation differences (cell width, leak policy)
that affect the ratio.
Three substantive findings:
1. Pure-compute parity with C: bench_compute_collatz runs at
AILang/C = 0.99x across both allocators. AILang's IR composes
with LLVM's optimizer at the same level a hand-C source does.
This is the LLVM-linkable performance claim, backed by data
for the first time. bench_compute_intsum (1.05-1.18x) confirms.
2. AILang bump beats glibc malloc 2x on linear allocation:
bench_list_sum.bump/c = 0.50x. Bump's two-instruction inline
fastpath outperforms glibc's free-list-managed malloc on
no-free workloads. Quantitatively measured for the first time.
3. RC overhead vs C malloc quantified: bench_list_sum.rc/c =
1.49x, bench_tree_walk.rc/c = 2.61x. The 8-byte refcount
header + zero-init + libc backing add 50-160% over glibc
malloc on these implicit-mode workloads. Explicit-mode + a
free()-adding C variant (21'f, queued) will close the
apples-to-apples gap on dec-cost.
CLAUDE.md updated to list bench/cross_lang.py as the third
tidy-iter gate alongside bench/check.py and bench/compile_check.py.
20 new metrics in bench/baseline_cross_lang.json with 12-15%
tolerances (cross-language ratios are inherently noisier than
within-AILang ratios — two compiler stacks contribute variance).
Closes the second axis the user named: every typechecker / codegen
perf change was previously invisible to the tidy-iter gate. With
Family 21 typeclasses (queued) and 21'b's poly-ADT additions both
pushing on the typechecker, naive substitution loops would have
landed silently and decayed the compile path.
bench/compile_check.py is a separate script from bench/check.py
because the methodology differs: sub-process spawn timing on small
workloads (1ms scale for `ail check`, 65ms for `ail build`) vs.
allocator-stress on large ones (multi-second). Tolerances differ
by an order of magnitude (25% / 20% here vs. 5-15% there).
Empirically: ail check is sub-ms across the corpus, dominated by
subprocess spawn (~5-10ms on Linux); ail build is 63-69ms,
dominated by clang's link step. The bench is a catastrophe
detector (10x slowdowns visible) — finer regressions need a
profiler. CLAUDE.md updated to list both scripts as co-equal
tidy-iter gates alongside the architect drift report; exit 0/1/2
semantics are uniform across both.
JOURNAL queue: 21'd (pure-compute fixtures) and 21'e (cross-
language reference / hand-C ratio) remain to land the LLVM-
linkable performance claim.
Adds a "Performance regressions" subsection to the iter-cycle
discipline. bench/check.py runs at every family close alongside
the architect drift report; its exit code is the gate. Skipping
it requires the same explicit JOURNAL entry as skipping the
drift review.
Codifies the discipline that 21'a's JOURNAL entry recommended
and explicitly punted to the orchestrator. The tooling shipped
yesterday; the rule that makes it load-bearing ships now.
Also clarifies two non-goals: per-metric tolerances are not the
language correctness bar (Decision-10 thresholds remain in
DESIGN.md, evaluated against absolute numbers); and improvements
do not auto-update the baseline — they surface for ratification.
The Telegram-via-PushNotification-tool path is replaced with a
direct ~/.claude/notify.sh "Text" invocation. Same payload
discipline (actionable one-liner, skip 'hi, I' framing); the
delivery channel is now under the user's control. Tested in the
previous session.
Bug fixes are red-then-green: write a failing test that pins the
symptom, make the smallest change that turns it green, keep the
test as a permanent regression. No fix without a test.
Bug fixes do not need orchestrator permission — when a bug is
unambiguous (build broken, fixture crashes, refcount underflow,
observable wrong output), the fix is autonomous orchestrator work.
The user directs features and priorities; bug fixes are mechanical
work that ships when the bug is found.
User wired up Telegram push notifications. New subsection under
"Direction freedom" pins down the discipline: only notify when
something is genuinely wanted from the user — exactly the
existing bounce-back triggers (real design judgement, genuinely
unexpected event, explicit checkpoint). Never for routine
progress or FYI updates.
The notification is the exception, not the rule. Default mode
stays autonomous. The notification body is the actionable
summary, one short line, phone-screen friendly.
Adds an "Iter cycle" top-level section: work clusters into named
iter families (18a–f, 19a–c, …); after each family closes, the
next iter is a non-optional tidy-iter that runs ailang-architect
over the surface and resolves every drift item by fixing,
ratifying in DESIGN.md, or recording acceptable drift in
JOURNAL.md. Skipping requires an explicit JOURNAL entry naming
the reason.
Also writes down the previously-implicit roles of DESIGN.md
(canonical spec; what AILang IS) and JOURNAL.md (decisions log;
HOW we got here, what's queued, rationale not in DESIGN.md). The
rule was practice for a stretch, then quietly evaporated when no
written rule kept it alive across sessions; codifying it here so
the next compaction doesn't lose it again.
User correction during the post-RC-overnight check-in: the 18a
"Type::Fn metadata vs. new Type variant" call was justified
across DESIGN.md, JOURNAL, and the commit message primarily by
"avoids ~250 match-arm sites". That is an observation about the
current state of the code, not a design rationale.
CLAUDE.md gains two binding rules:
- Design rationale ≠ implementation effort. Effort is at most a
tiebreaker; a choice whose only stated reason is effort is
suspect. The rule names the 18a misstep as the canonical
anti-example so future sessions catch it earlier.
- Direction freedom + bounce-back conditions inlined (was
previously a cross-reference to a private auto-memory file
outside the repo, which the user couldn't see).
DESIGN.md Decision 10's Schema-additions block now leads with
the substantive reasons for per-position metadata: semantic
locality (modes belong to fn-parameter positions, not to types
in general — Decision 1 line); compositional clarity (type
identity vs. calling convention factor apart); future-proofing
(per-position metadata generalises; Type-variant approach
combinatoric blows up). The match-arm count remains parenthetical,
explicitly named a tiebreaker.
JOURNAL records the correction itself — the mistake stays as a
data point because how design discipline corrupts is informative.
Adds an explicit "My role: orchestrator" section so the discipline
doesn't drift across sessions. Three concrete contracts:
- What I do myself (planning, design, JOURNAL/DESIGN) vs. what gets
delegated (implementation, testing, debugging, architectural
drift review). Trivial mechanical edits stay inline.
- Authority over /agents/: I may add, edit, retire, or replace agent
definitions when orchestration needs change, with the change going
through git like any other code.
- When NOT to delegate: direct user questions, single judgement
calls, and cases where context is already loaded.
Triggered by a session where I quietly took over implementer work
on Iter 13a instead of routing it through ailang-implementer.
Finishes the project-wide English convention:
- CLAUDE.md (the user opted in to translate it too).
- examples/hello.ail.json: "Hallo, AILang." -> "Hello, AILang.".
- E2E test assertion and IR snapshot for hello regenerated to match.
- DESIGN.md "Project language: English" section: the previous
CLAUDE.md exception is dropped.
After this commit, no German remains in any committed file
(grep -P '[äöüÄÖÜß]' is empty across .rs / .md / .json / .ll).
Verified: cargo test --workspace passes (44/44).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Erste lauffähige Iteration. examples/sum.ail.json wird zu nativem Binary
kompiliert und druckt 55 (Summe 1..10) als End-to-End-Test.
Architektur:
- ailang-core: hashbares JSON-AST + canonical-form + pretty-printer
- ailang-check: monomorpher HM-Subset + Effekt-Set-Tracking
- ailang-codegen: LLVM-IR-Text-Emitter (kein libllvm-link)
- ail: CLI mit check/manifest/render/describe/emit-ir/build/builtins
Designentscheidungen sind in docs/DESIGN.md dokumentiert; der Verlauf
in docs/JOURNAL.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Brummel