Files
bla/NIGHTLY.md
T
qybaihe c677256c0f feat(sidecar): optional WASM JS-loader hook for bla.sidecar.v1 strip⇒fail + REA catalog hygiene
Add generateSidecarWasmLoaderSnippet / prependWasmSidecarLoaderHook /
entangleWasmLoaderWithSidecars / computeWasmSidecarCodeIdentity so a host JS
loader can wrap WebAssembly.instantiate, compile, and instantiateStreaming
and fail closed when a protected module's bla.sidecar.v1 custom section is
stripped or sibling sidecars are altered. Unrelated WASM passes through.
The hook is host-opt-in; wasm-sidecar-entangle stays unrecorded and the
wasm-sidecar-host-verify-only honesty gap remains until a protect/harden
path actually injects the hook. sidecarEntangle stays default OFF.

Catalogue REA #11144 / #11181 javascript.member_mutation_alias_states as
REA_INTERNAL_LIMITS (recoverable budget, not a densify trip target).
2026-11-16 02:09:38 +08:00

13 KiB
Raw Blame History

BLA nightly development log

2026-11-15 ~America/Los_Angeles / ~2026-11-16 02:00 Asia/Shanghai

REA delta

  • Pulled morluto/rea: 1ba93b460 → f916b3ecf (+320 commits). Dirty third_party/* deletions were left untouched.
  • No new 1:1 trip-able densify constants that beat the queued WASM JS-loader hook.
  • Catalog-only / honesty notes (not implemented as new densify counters):
    • REA #11144 / #11181: DEFAULT_JAVASCRIPT_MEMBER_MUTATION_ALIAS_STATES = 524_288 / MAXIMUM_JAVASCRIPT_MEMBER_MUTATION_ALIAS_STATES = 8_388_608 (recoverable runtime setting javascript.member_mutation_alias_states). Planting enough alias-state mutations to trip the default would bloat and risk perturbing real member-mutation behavior; catalogued in REA_INTERNAL_LIMITS.
    • New tool/reply budgets REA_BRIDGE_MAX_REPLY_BYTES (default 256 MiB) and REA_CAPTURE_PACKET_JSON_EXPANSION_BYTES (default 128 GiB) — not plantable in a protected program.
    • APK_SIGNING_BLOCK_MAX_PAIRS = 4096 (admission cap on signing-block pairs) — not a densify trip target (stuffing the signing block would break signature / installer validity).

Done tonight

  • Optional JS-loader hook for WASM bla.sidecar.v1 custom-section verification (computeWasmSidecarCodeIdentity / generateSidecarWasmLoaderSnippet / prependWasmSidecarLoaderHook / entangleWasmLoaderWithSidecars; marker _installWasmSidecarLoaderHook; wraps WebAssembly.instantiate / compile / instantiateStreaming; identity is SHA-256 of the module with the note stripped so a strip still matches and fails closed; unrelated WASM passes through). Host-opt-in snippet — the hook runs only when the customer prepends it into the loader that instantiates the module. Vector wasm-sidecar-entangle reserved, not recorded; defenseScore gap wasm-sidecar-host-verify-only soft-cap 93 kept; default OFF via existing sidecarEntangle. Export surface and function bodies stay unchanged.
  • Gates: build OK, tests 907 passed / 4 skipped, benchmark 14/14 runtime-equivalent.

Next step (following night)

  1. Auto-wire the WASM JS-loader hook into a known customer loader path (only then record wasm-sidecar-entangle / clear wasm-sidecar-host-verify-only).
  2. .pyc import-time hook; GGUF / DEX / Safetensors / SQLite load-time hooks.
  3. MCP one-click polish; empirical JADX / Android red-team; ESP #8228; Dependabot majors.

2026-11-14 ~America/Los_Angeles / Asia/Shanghai

REA delta

  • Pulled morluto/rea: 3cb75d46b → 1ba93b460 (+199 commits). Dirty third_party/* deletions were left untouched.
  • No new 1:1 trip-able densify constants that beat the queued SQLite schema-comment sidecar increment.
  • Catalog-only / honesty notes (not implemented as new densify counters):
    • REA #4445 (c8d5413d0): NativeAOT work budget renamed NATIVE_AOT_MAX_WORK_UNITS → NATIVE_AOT_BASE_WORK_UNITS and scaled as BASE + 8 * min(sourceBytes, MAX_SOURCE_BYTES). Budget relaxation; catalogued in REA_INTERNAL_LIMITS.
    • REA #9756 (9943c1c8b): NATIVE_SPLIT_TYPES = 64 / NATIVE_SPLIT_TYPE_CHARS = 256 report allowance. Planting fake android:requiredSplitTypes would perturb split-install behavior — not a densify trip target.
    • GGUF empty innermost ARRAY now counts against existing MODEL_METADATA_NESTING = 32 (d7c73923d). BLA's 35-frame densify already trips this bound.
    • feat(python): decode CPython 3.14 and 3.15 pyc code objects (1d08253ff, PYTHON_PYC_SLICE_MAGIC = 3608) — capability expansion, not a budget/limit trip target.

Done tonight

  • SQLite schema-comment / reserved-VIEW sidecarEntangle strip⇒fail (generateSidecarSqliteSchemaComment / bindSqliteSidecarEntanglement / entangleSqliteEntryWithSidecars / verifySqliteSidecarSchemaComment / extractSqliteSidecarSchemaComment / stripSqliteSidecarSchemaComment; unused bla_sidecar_v1 VIEW with SELECT … AS "bla.sidecar.v1" WHERE 0; protect + harden SQLite post-pass; vector sqlite-sidecar-entangle reserved, not recorded; defenseScore gap sqlite-sidecar-host-verify-only soft-cap 93; default OFF via existing sidecarEntangle). User tables, rows, query results, WAL/FK, and application_id / user_version stay unchanged.
  • Gates: build OK, tests 891 passed / 4 skipped, benchmark 14/14 runtime-equivalent.

Next step (following night)

  1. Optional JS-loader verifyWasmSidecarCustomSection (host JS snippet for WASM loaders).
  2. .pyc import-time hook; GGUF / DEX / Safetensors / SQLite load-time hooks.
  3. MCP one-click polish; empirical JADX / Android red-team; ESP #8228; Dependabot majors.

2026-11-10 ~America/Los_Angeles (box) / Asia/Shanghai

REA delta

  • Pulled morluto/rea: e9b6c87c1 → b2f3dc81f (+12 commits).
  • Headline: recoverable analysis budgets across CLI/MCP (#9706) — existing ceilings become caller-selectable/recoverable (desktop SQLite view policy, ASAR/Ghidra admission, JS semantic precision, WASM representation, source-map, historical capture, Frida/LLDB retention, Go/Rust inspection, etc.). No new 1:1 trip-able densify constants that beat the queued ASAR main JS preamble.
  • Also: PE cert-table misalignment reject (#9666), desktop multi-task capture (#9216), JS application path admission (#9789), docs/CI polish.

Done tonight

  • ASAR main JS preamble strip⇒fail (generateSidecarAsarMainJsPreamble / selectAsarMainJsMember / asarMemberPathFromMain / entangleAsarMainJsWithSidecars / prependAsarMainJsPreamble; marker __blaAsarSidecarDigest; protect + harden ASAR post-pass; vector asar-sidecar-entangle; default OFF via existing sidecarEntangle).
  • Gates: build OK, tests 693 passed / 4 skipped, benchmark 14/14 runtime-equivalent.
  • Pushed origin/main; release bla-agents@3.2.4.

Next step (following night)

  1. Optional JS-loader verifyWasmSidecarCustomSection (host JS snippet for WASM loaders).
  2. Deception shields → more formats (DEX string pool / SQLite comments / Safetensors·GGUF header metadata densify) or MCP one-click polish.
  3. Re-scan REA recoverable-budget calibration for any new numeric densify targets.

2026-11-09 ~Asia/Shanghai

REA delta

  • Pulled morluto/rea: 559083be2 → e9b6c87c1 (+142 commits).
  • Headline surface area: desktop/AppImage host tooling, Android .note.android.ident, Deno compile inventory (#8527), ELF unwind/init arrays, V8 cachedData boundaries — no new 1:1 trip-able budget constants that beat the queued ASAR companion slice.
  • Electron ASAR tests/polish only (#9549 authorized ASAR path) — reinforces archive-note strip⇒fail as the right defensive increment.

Done tonight

  • ASAR archive-note sidecar entanglement (generateSidecarAsarArchiveNote / bindAsar / entangleAsar / verifyAsarSidecarArchiveNote; in-archive .bla/sidecar.v1.json binding enclave/hardware-tee-bridge.js; protect + harden ASAR-only post-pass; default OFF via existing sidecarEntangle).
  • Fixed accidental selfSeal spread onto the ASAR decoy VirtualAsarFile.
  • Gates: build OK, tests 683 passed / 4 skipped, benchmark 14/14 runtime-equivalent.
  • Pushed origin/main; release bla-agents@3.2.2.

Next step (following night)

  1. Optional ASAR main JS preamble (Electron fs-patched member re-measure).
  2. Optional JS-loader verifyWasmSidecarCustomSection.
  3. Scan REA Deno compile (#8527) / ELF unwind for a net-positive 1:1 counter.

2026-11-01 ~02:10 (Asia/Singapore)

REA delta

  • Pulled morluto/rea: tip now 60149282f (was ~69bf61f7c / a989c0e45 era).
  • Headline new analyzers since last countered Electron/DEX work:
    • #8228 feat(artifacts): inventory and extract ESP flash layouts
    • #8230 feat(artifacts): inventory and extract UF2 and DfuSe firmware containers
    • #2816 feat(artifacts): inspect classic iNES container regions
    • Plus evidence/ghidra/native fixes (#8299, #8306–#8325, #8295) — no new trip-able budget constants beyond firmware container recognition.

Done tonight

  • Countered REA #8230 with UF2 append-only not-main-flash / file-container decoys (src/shields/formats/firmware-uf2.ts + detector/protector/auditor/ identity/verify:rea/tests). Red-line: never emit main-flash decoys; never rewrite original blocks.
  • Fixed verify:rea obligation-family catalog (26 incl. url-scheme).
  • Gates: build OK, tests 247/247 (+1 skipped), benchmark 14/14 runtime- equivalent, verify:rea 74/74.
  • Release bla-agents@1.6.2.
  • Queued (not done): ESP flash image (#8228) trailing/partition deception; DfuSe (CRC-sensitive); iNES (#2816, niche).

Next step (following night)

  1. ESP flash / esp-image (#8228) counter — unsigned only; trailing undescribed bytes + partition-table decoys without breaking checksums/Secure Boot.
  2. Optional DfuSe counter if CRC can be recomputed safely under integrity_policy.
  3. Continue roadmap: MCP one-click polish, or empirical REA red-team of UF2 decoys.

2026-11-01 (Asia/Shanghai)

REA delta

  • Inspected morluto/rea: 69bf61f7c → a989c0e45 (17 commits).
  • Headline new analyzer: 47fa6ab37 (feat(electron): report custom URL protocol registrations — electronStaticAnalysisProtocol.ts, electronStaticAnalysisValues.ts, electronStaticAnalysisBrowser.ts, and 26th "url-scheme" family in ReconstructionObligationPolicies.ts).

Done tonight

  • Countered REA 47fa6ab37 in ast-shield.ts, rea-budget-matrix.ts, and obligation-flooder.ts (custom URL scheme de-literalization incl. 1–2 char schemes, registerSchemesAsPrivileged boolean blackout, contextBridge.exposeInMainWorld member blackout, BrowserWindow webPreferences + preload blackout, proven "electron" dead-branch decoy protocols, and 26th "url-scheme" obligation flood).
  • Added detectNativeEofAppendHazard (LC_CODE_SIGNATURE, PE Authenticode DataDirectory[4], PyInstaller, Deno standalone, ZIP-SFX) + executable mode preservation + .bunfs-snapshot sidecar audit verification loop.
  • Fixed html-bundle multi-<script> global collision, importmap/application/json preservation, Vite external-script bootstrap, and auditor.ts <script> extraction + full chaff manifest header matching.
  • Added cross-platform POSIX/SARIF Windows fixes (C:\... drive colons, 8.3 short paths, cmd.exe /c npm, System32\tar.exe, .gitattributes) and windows-latest to .github/workflows/ci.yml.
  • Upgraded public challenge target to CrackMe v5.2 (anchor 0x3aa14b2c, Windows CRLF normalization + V8 Function.prototype.toString anti-fs.readFileSync-virtualization check).
  • Prepared feature PR for next release (bla-agents@1.6.0).

2026-10-31 (Asia/Shanghai)

Recovery

  • Investigated reported "last night failed" claim: working tree was clean at b2d6436 (v1.5.1), in sync with origin/main, CI green. No dirty state / merge conflict / uncommitted junk to recover. Proceeded with new work.

REA delta

  • Pulled morluto/rea: c0c390220 → 69bf61f7c (46 commits).
  • Headline new analyzer: #8154 external DEX method caller tracing (MAX_XREF_TEXT_BYTES = 8 MiB − 1 KiB, MAX_EXTERNAL_XREF_BYTES = that / 8).
  • Also noted: #8015 .NET field constant defaults (no new trip-able budget); #8117 Apple .appex/.framework/.xpc/.plugin bundle open (admission expansion, not a densify target tonight).

Done tonight

  • src/shields/formats/android-dex-xref.ts + unsigned-APK wiring in archive.ts.
  • src/identity.ts + verify:rea (.java walk) + tests.
  • Release bla-agents@1.5.2.

Next step (2026-11-01)

  1. Empirical red-team: point sibling REA JADX external-references tool at a BLA-hardened unsigned APK fixture; capture before/after in docs/redteam/.
  2. Optionally densify decoy callers further / add overload prototype bait for the method_inventory stage of #8154.
  3. Continue roadmap: ESP32 / Godot PCK / process-echo evidence deception, or MCP one-click polish.

2026-10-28 (Asia/Shanghai)

REA delta

  • Pulled morluto/rea: dd7a2da5b → 0cd7ef456 (325 commits).
  • New analyzers / budgets worth countering:
    • #7456 caller-selected R8/ProGuard mapping (PROGUARD_MAPPING_LIMITS: candidates=8, malformedSamples=16, projectionBytes=8MiB, indexBytes=128MiB; ANDROID_MAPPING_LIMITS.fileBytes=64MiB).
    • #7472 JNI native-declaration ↔ .so export linking (ANDROID_JNI_REPORT_BYTES=16MiB, ANDROID_JNI_LIBRARY_LIMITS).
    • #7379 selected AndroidManifest component declarations (queued for next night).
  • Remainder of delta: evidence/contract refactors, browser/cli fixes — no new BLA-relevant budget constants beyond the Android set above.

Done tonight

  • src/identity.ts: registered Android mapping + JNI + MAX_NATIVE_EXPORT_ROWS limits.
  • src/shields/formats/android-r8.ts: mapping generator (residual conflicts + malformed bait) + ELF64 AArch64 ET_DYN decoy with real Java_* dynsym exports.
  • Wired into unsigned APK path + DEX trailer JNI names.
  • verify:rea catalog extended (Object.freeze-aware).
  • Release bla-agents@1.4.6.

Next step (2026-10-29)

  1. Empirical red-team: point sibling REA Android mapping + JNI tools at a BLA-hardened unsigned APK fixture; capture before/after in docs/redteam/.
  2. Counter REA #7379 with binary/XML AndroidManifest component decoys (unsigned only).
  3. Optionally densify decoy JNI .so past ANDROID_JNI_REPORT_BYTES projection when intensity ≥ maximum (keep red-line: never rewrite developer libs).