SeymourBETA

Aggiornamento 2.195

31/07/2026

v2.195.0.0BETAMINOR

Seymour31 lug 2026

P2-B L2 (LLM contract HTTP adapter) + P2 L1 (ATEL semantic IR + three-layer diff)

LLM contract L2 (Core/LLM/LlmHttpAdapter.cs + 8 tests): model call with timeout (CancelAfter), cancellation, MaxResponseBytes bounded stream read, strict JSON parse (JsonUnmappedMemberHandling.Disallow — unknown field = MalformedPayload), key ONLY in Authorization: Bearer header (never body/log/exception; pattern redaction + exact session-key redaction), non-loopback host refuses without confirmation BEFORE touching the network, base64 validated, failures always become typed LlmResult with redacted message (no raw stack). LlmCredential gained a controlled internal accessor (ReadApiKeyCopy — copy, never serialized; reflection removed). Zero real network in tests (fake DelegatingHandler).
P2 L1 — ATEL semantic IR (FfxLib/Ai/AiSemanticScript.cs + AiDiffThreeLayer.cs + 10 tests): 100% additive overlay over the codec (AiScript_File fully untouched, RT0 361/361 preserved): AiEvidenceState (Decoded/Hypothesis/Confirmed — L1 produces only Decoded), AiInstructionNode per (worker, offset) with Meaning enriched via the real AiCommandId.Decode (e.g. PUSHII -> Firaga); ThreeLayerDiff = bytes layer + disassembly (AiScript_Diff.Compare reused) + semantics (command names/evidence). Validated against real fixtures m000.bin/m001.bin (real AiFile from corpus) + byte-identical synthetic RT0 fixture. Subagent expectation fix: m000's 3C D8 5F 00 3C 3C script has 4 instructions (RET | CALLPOPA 3B | RET | RET), not 3 — Build was right.
Baseline: 185/185 PASS (164 → 185). ### MINOR

v2.194.0.0

Torna agli aggiornamenti

SeymourBETA

Aggiornamento 2.195

31/07/2026

v2.195.0.0BETAMINOR

Seymour31 lug 2026

P2-B L2 (LLM contract HTTP adapter) + P2 L1 (ATEL semantic IR + three-layer diff)

LLM contract L2 (Core/LLM/LlmHttpAdapter.cs + 8 tests): model call with timeout (CancelAfter), cancellation, MaxResponseBytes bounded stream read, strict JSON parse (JsonUnmappedMemberHandling.Disallow — unknown field = MalformedPayload), key ONLY in Authorization: Bearer header (never body/log/exception; pattern redaction + exact session-key redaction), non-loopback host refuses without confirmation BEFORE touching the network, base64 validated, failures always become typed LlmResult with redacted message (no raw stack). LlmCredential gained a controlled internal accessor (ReadApiKeyCopy — copy, never serialized; reflection removed). Zero real network in tests (fake DelegatingHandler).
P2 L1 — ATEL semantic IR (FfxLib/Ai/AiSemanticScript.cs + AiDiffThreeLayer.cs + 10 tests): 100% additive overlay over the codec (AiScript_File fully untouched, RT0 361/361 preserved): AiEvidenceState (Decoded/Hypothesis/Confirmed — L1 produces only Decoded), AiInstructionNode per (worker, offset) with Meaning enriched via the real AiCommandId.Decode (e.g. PUSHII -> Firaga); ThreeLayerDiff = bytes layer + disassembly (AiScript_Diff.Compare reused) + semantics (command names/evidence). Validated against real fixtures m000.bin/m001.bin (real AiFile from corpus) + byte-identical synthetic RT0 fixture. Subagent expectation fix: m000's 3C D8 5F 00 3C 3C script has 4 instructions (RET | CALLPOPA 3B | RET | RET), not 3 — Build was right.
Baseline: 185/185 PASS (164 → 185). ### MINOR

v2.194.0.0