这是一次单轮规格驱动的自主端到端工程实验：人只输入了一次任务规格， AI Agent（Claude Code + 自研 plan-execute-audit harness）独立完成目标理解、 技术选型、规划、架构、实现、验证、CI、外部审查修复与合并。全程技术指导为 0、人工代码修改为 0 —— 共 3 次人类介入，全部为权限/环境类，逐条记录于 intervention-log。
工程成果：Rust 从零实现的 WebAssembly 2.0 解释器，以官方 test/core 全量 （wg-2.0 冻结快照，148 个测试文件）为验收语料：54,006 条命令全部判定， 52,915 PASS / 0 FAIL / 1,091 UNSUPPORTED（后者全部是文本格式方法边界， 非语义豁免），并与同一 commit 构建的官方 OCaml 参照解释器交叉验证。
自主实验过程：从规格输入（2026-07-17 16:01:48Z）到 PR 合并 （2026-07-18 00:53:39Z）约 9 小时；期间通过 6 轮外部 Codex review （29 项发现全部修复）、独立计划/执行审计、3 组负向对照与 fresh-checkout 复现（25 秒全管线）。
深入阅读：中文 case study · 证据直达链接 · 最短复现路径
A from-scratch WebAssembly 2.0 interpreter in Rust, acceptance-tested
against the complete official test/core suite at the W3C wg-2.0
snapshot with fail-closed accounting, and cross-checked against the
official OCaml reference interpreter built from the same commit.
This repo is simultaneously two deliverables:
- The engineering result — the interpreter itself and its frozen, CI-enforced acceptance (tables below).
- An autonomous-agent experiment — the entire implementation was
produced in a single-spec autonomous run: one human-written
specification in, merged and reproducible repo out, with zero
technical guidance and zero hand-written code. Counted human
interventions: 3, all permission/environment unlocks; the log also
discloses one non-counted exchange (the user relayed public
review-bot comments, which the agent re-fetched independently).
See Provenance and the
frozen run records under
goal-runs/wasm-zero-oneshot/.
| Metric | Value |
|---|---|
| Corpus | WebAssembly/spec tag wg-2.0 (fffc6e12), test/core 90 files + test/core/simd 58 files |
| Total commands | 54,006 |
| PASS | 52,915 |
| FAIL | 0 |
| UNSUPPORTED | 1,091 (all module_type == "text"; see method boundary below) |
Accounting is no-silent-skip: every command receives exactly one
verdict in {PASS, FAIL, UNSUPPORTED}, the runner's ledger is
cross-checked by an independent enumeration script
(scripts/enum_corpus.py), and unknown command/action/value shapes are
FAIL, never skipped.
The repo was built end-to-end by an AI agent (Claude Code driving a
plan-execute-audit harness, recorded as runner: claude-gah) from one
human-written specification. It is an experiment in spec-driven
autonomous engineering — not "one sentence in, code out": the input was
a precise acceptance contract, and everything downstream (planning,
architecture, implementation, oracle setup, negative controls, CI, code
review remediation, merge) was done by the agent.
| Autonomy metric | Value | Evidence |
|---|---|---|
| Human-written specification | 1 (the initial prompt; frozen as AC1-AC7) | contract.md |
| Technical guidance / hand-written code | 0 / 0 | intervention-log.md |
| Human interventions (all permission/environment) | 3 | same log, per-entry classification |
| Wall clock, spec input → merged PR | ≈ 9 h (2026-07-17 16:01:48Z → 2026-07-18 00:53:39Z) | contract.md L68; PR #1 |
| External review | 6 Codex rounds, 29 findings (7 P1, 22 P2), all fixed; round 6 clean | PR #1; fix commits 2a4fc2d ba95121 d3f8d67 44df91d 9a0622f |
| Negative controls (accounting must turn red) | engine defect → FAIL=328; validator accept-all → FAIL=2,146; unknown inputs → 4/4 FAIL | negctl logs |
| Fresh-checkout reproduction | full pipeline 25 s, exit 0 | evidence.md |
flowchart LR
A["Single spec input<br/>(frozen contract AC1-AC7)"] --> B["Plan +<br/>independent plan audit"]
B --> C["Corpus + oracle baseline<br/>148 .wast via wast2json;<br/>OCaml ref interpreter 148/148"]
C --> D["Implementation<br/>decoder / validator /<br/>interpreter / SIMD"]
D --> E["Fail-closed runner +<br/>double-entry accounting"]
E --> F["Negative controls A/B/C"]
F --> G["CI hard gate:<br/>full 54,006 sweep"]
G --> H["Independent<br/>execution audit"]
H --> I["PR #1 +<br/>6 Codex review rounds"]
I --> J["Merge 3d06be1"]
J --> K["Fresh-checkout<br/>reproduction, 25 s"]
| Claim | Where to verify |
|---|---|
| Frozen acceptance contract (AC1-AC7, written before implementation) | contract.md |
Final ledger: 54,006 / 52,915 / 0 / 1,091, accounting_ok |
ledger-final.json (totals block at end) |
| Every human intervention, classified | intervention-log.md |
| Timeline and milestone log | execution-log.md |
| Negative controls A/B/C raw output | negctl-a · negctl-b · negctl-c |
| Oracle sweep (OCaml reference, 148/148 exit 0) | oracle-sweep.log |
| External review thread (6 rounds, 29 findings) | PR #1 |
| CI definition + latest runs | ci.yml · Actions |
| 深度中文解读（任务难点、失败与恢复、review 价值、non-claims） | docs/case-study.zh-CN.md |
crates/wasm-core— the engine (zero dependencies):decode.rs— strict wg-2.0 binary decoder (LEB128 width/sign checks, section order/size enforcement, full opcode space incl.0xFCand the SIMD0xFDspace)validate.rs— the spec-appendix type checker (operand/control stacks with bottom values), constant-expression rules, declared function references, import/limit checksstore.rs/exec.rs/simd.rs— instantiation per wg-2.0 semantics and an iterative interpreter (explicit frame stack, so call-stack exhaustion is a deterministic trap); floats are carried as raw bits end-to-end to preserve NaN payloads
crates/spec-runner— drives the wast2json-converted corpus and produces the ledger. Judgement rules live here (NaN patterns, trap-message matching, malformed-vs-invalid stage separation).
The implementation surface is the binary format. The text-format
frontend is delegated to the official WABT toolchain (wast2json
1.0.41, SHA256-pinned): .wast scripts are converted to JSON commands
plus binary modules. The 1,091 assert_malformed assertions whose
module is given as quoted text target the text parser itself and are
counted UNSUPPORTED with per-row attribution in the ledger — this is a
method boundary, not a semantic exemption. All 2,146 assert_invalid
and 719 binary assert_malformed assertions are in the driven set and
pass.
Trap-message matching rule: an assert_trap passes iff the action traps
and one message is a prefix of the other (spec-canonical messages, e.g.
uninitialized element vs uninitialized element 7). Invalid/malformed
expectations require the correct failure stage (decode vs validation),
not message equality.
Prerequisites: Rust (stable), Python 3, curl, git.
scripts/fetch_spec.sh # clone spec pinned at wg-2.0 (fffc6e12)
scripts/fetch_wabt.sh # WABT 1.0.41 release, SHA256-verified
scripts/convert_corpus.sh # 148 .wast -> build/wast-json/
cargo run --release -p spec-runner -- \
--expect-total 54006 --expect-unsupported 1091 \
--ledger-out build/ledger.json
python scripts/enum_corpus.py --ledger build/ledger.jsonBoth commands exit nonzero on any FAIL or accounting mismatch. CI runs
exactly this sequence plus cargo fmt --check, clippy -D warnings,
and unit tests.
The corpus-oracle consistency proof runs the official OCaml reference
interpreter — built from the same spec commit — over all 148 .wast
files (each exits 0). This is not part of the CI gate; to re-run it:
# any Linux/WSL with opam >= 2:
opam switch create wasm-zero ocaml-base-compiler.5.2.1
eval $(opam env --switch=wasm-zero)
opam install -y dune menhir
cd third_party/spec/interpreter && make # produces ./wasm
cd .. && for f in test/core/*.wast test/core/simd/*.wast; do
./interpreter/wasm "$f" || echo "ORACLE_FAIL: $f"
doneDuring development, disputed decode/validate classifications were
adjudicated against the suite + oracle (never by editing expectations);
the three adjudicated nuances are documented in decode.rs comments and
goal-runs/wasm-zero-oneshot/evidence.md.
- Implements WebAssembly 2.0 (wg-2.0 snapshot): SIMD, bulk memory, reference types, multi-value, sign-extension, saturating truncation, multiple tables, import/export of mutable globals.
- Not implemented: Wasm 3.0 features (GC, exception handling, memory64, multi-memory, tail calls, relaxed SIMD), the text format (delegated to WABT), threads, and any embedder API beyond what the spec-runner needs.
- This is a correctness-first interpreter, not a sandbox hardened for untrusted production workloads, and not performance-tuned.
- "Full test/core" always means the wg-2.0 snapshot of the suite.
Apache-2.0.