Code review after the initial prove commit found and fixed:
1. Calls inside table/list literals were silently dropped — a function
with side-effectful cells lowered to "no ops" and claimed purity in
Lean. lower_expr's fallback now rejects with proof.unsupported
(author must bind the call to a variable first).
2. Same-callee call sites with different arities (optional params)
shared one Op whose ctor declared the first-seen arity, emitting
uncompilable Lean. register_op now suffixes `/N` per arity.
3. proof.name_shadow fired once per walked function (N functions → N
duplicates) and checked params of claim-free functions. Predicate
check runs once per module; param check only for functions that
carry claims.
4. 长度(轨迹) = 2 emitted mqEq (Nat) (Val) — type-mismatched Lean.
Nat-context detection renders trace-length comparisons natively;
untranslatable shapes become comments instead of broken Lean.
Hardening: lean_ident reserves 轨迹₀ / h_前提 (theorem binder
collision); emit_theorems dedups repeated claim ids.
Tests: +4 integration regressions (p2_a..p2_d), 23 prove tests total;
full suite green except the documented env-limited failures (one
host::foreign_session flake passes standalone).
Business validation (doc §12): e-commerce order-approval SKILL with 3
theorems cross-checked against the real interpreter — 5000 → auto
(1-step trace), 20000/60000 → manual ticket T-20000/T-60000
(需人工→创建, exactly 2 steps); corrupted claim tables hit
check --proof codes while staying default-off; double emit
byte-identical.
Send complete op/steps/value records to the Lean decoder for every operation. Preserve companion exit status and a bounded stderr tail on transport failure. Register the smoke fixture read tool explicitly without weakening the allowlist or changing its output assertion.
Validation: 462 library tests and 312 gold tests passed. The repository regression run passed 946 tests with two existing ignored experiments; the previously passed 1000-document scale test was not repeated.
Add root/index and layout skills that force file-partitioned pages and ban
HTML/CSS injection as the default path; wire ext add, docs, and call coverage.
Co-authored-by: Cursor <cursoragent@cursor.com>
Root skills get a choose-a-path table; circuit modules spell the simulate
handoff; call_cli covers ZH bell and stepwise session probabilities.
Co-authored-by: Cursor <cursoragent@cursor.com>
Split ext/quantum into index + modules, add marqdo call --effects/--session/--toget,
narrative-ratio checks, and a same-package 量子/ tree so agents can follow a
decision path for Bell demos without inventing a second API.
Co-authored-by: Cursor <cursoragent@cursor.com>
Host exposes oracle journals as language values with refuse/copy gates; bytecode and foreign session paths get modularized; applications and verification replay-policy sources land with Lake builds ignored.
Co-authored-by: Cursor <cursoragent@cursor.com>
The same page is the document, the knowledge and the program. marqdo call runs that function with arguments, so reuse, tools and tests do not need a second skill system.
Co-authored-by: Cursor <cursoragent@cursor.com>
Env port strings must be json.parsed to int; pin main site to 18220 so
inherited MARQDO_SITE_PORT cannot steal the fixture bind.
Co-authored-by: Cursor <cursoragent@cursor.com>
Serve static application/json and text/csv previews with outline/table
limits and flip extract_formats availability for those media types.
Co-authored-by: Cursor <cursoragent@cursor.com>
Prepare/start/commit write and execute tools with idempotent dedupe, and refuse
blind retry while ledger state is unknown or interrupted mid-flight (MA-T03).
Co-authored-by: Cursor <cursoragent@cursor.com>
Open AgentOutputV1/ContextPackV2 capability bits and add EffectLedgerRecordV1
so CheckpointV2 write auto-resume only unlocks on a safe watermark.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add ToolSpecV1 register/search/describe/prepare so CALL and native paths share one schema gate; unauthorized and invalid-arg prepares fail with zero executions.
Co-authored-by: Cursor <cursoragent@cursor.com>
Replace bare resume writes with temp/fsync/rename, mark legacy loads as manual_review, and expose bounded read-only retry decisions while keeping write auto-resume closed without a ledger.
Co-authored-by: Cursor <cursoragent@cursor.com>
Stop full recompile on every knowledge preflight by writing ekc/graph/ir.json
and loading it unless --refresh or missing projections, cutting chat预检 wall time.
Co-authored-by: Cursor <cursoragent@cursor.com>
Observer SSE close no longer cancels execute by default; short/read/control
bypass parent INVOKE_LOCK via isolated admission, with read_lane=1 and gold samples.
Co-authored-by: Cursor <cursoragent@cursor.com>
Hold analysis locks across compile for concurrent same-snapshot reads,
emit SnapshotManifestV1 for app retention, and add Go drainForward tests.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add dual-endpoint LLM isolation, usage revision replace, cancel/F08/F11
asserts, and mark the design status Partial until remaining F gaps close.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add capabilities/selection/usage/DocumentView/locator/RunContext with
honest probes, LLM messages+cancel, sse-live cancel bridge, and CI
workbench_contracts gate; record design evidence for verified slices.
Co-authored-by: Cursor <cursoragent@cursor.com>
Make examples/executable-knowledge runnable on both backends, bind same-module
Type.member paths, add closure digests to locks/impact, and gate 100 corpus
units plus CLI smoke so S4–S6 leftover items have reviewable evidence.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add artifact.extract_pure for restricted expression lifting, a runnable
bookstore tax story fixture, and G12 discovery limits so oversized or deep
trees fail with explicit diagnostics instead of hanging.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add comment-line NarrativeRef steps, generation CURRENT snapshots, frontmatter
sources/applicability, digest-guarded artifact.patch, lock change classes, and
an offline CI gate so the documented S0–S6 path has reviewable evidence.
Co-authored-by: Cursor <cursoragent@cursor.com>
Share resolve_for_exec across Tree/Bytecode and semantic projections,
emit bound_calls/dynamic_calls, and require eng_reuse bind_targets before
execute so discovery scores cannot be treated as handles.
Co-authored-by: Cursor <cursoragent@cursor.com>
Imported functions shared the flat call table, so bare same-named calls
could silently bind across imports. Scope tags keep path calls working
while rejecting ambiguous bare names on both backends.
Co-authored-by: Cursor <cursoragent@cursor.com>
Establish author baseline fixtures and offline tests, make SemanticProject
the sole authority for graph edges, classify named steps, and tighten
local reuse so unknown effects are not pure and MLSP binds by SymbolId.
Co-authored-by: Cursor <cursoragent@cursor.com>