coreci/nextcraft
3
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 24 Wiki Activity

24 Releases 30 Tags

RSS Feed
  • v0.2.8 e798e1a6da
    Compare

    v0.2.8 — Nextcraft v0.3 Credential Engines (milestone release) Stable

    coreci-bot released this 2026-09-12 21:44:58 +00:00 | 6 commits to main since this release

    Nextcraft v0.3 — Credential Engines

    The six AI tutor agents now operate on REAL credential engines. v0.2's mock-input caveat is retired: Lab/Assessor/Proctor consume live telemetry, stored grades, and defense integrity signals — verified with zero corpus fixtures in the learner path.

    Shipped

    • Sandbox fabric (REQ-3-001/002): per-learner Linux-namespace sandboxes (unshare user/mount/pid/net — probe-verified isolation), lifecycle API, orphan reaper, G-5 abuse control (allowlist + per-learner/global caps) in place of auth
    • Live build telemetry (REQ-3-003): stdlib-only in-sandbox capture agent, WebSocket ingest with (learner,task,seq) dedup, gap detection, 1008 flood close + INCOMPLETE_FLOODED (G-3); at-least-once delivery / exactly-once storage proven end-to-end; SQLite (WAL) first real persistence
    • Process-trace grading (REQ-3-004): deterministic digests (D-028 — LLM sees digest only), rubric via the D-020 defense, BINDING G-4 gate (gapped/flooded traces are UNGRADABLE — no credential from an incomplete trace), calibration contract over D-021-aligned archetypes
    • Per-learner variants (REQ-3-005): sha256(template|learner|milestone) seeds, typed slot sampling, cache-first (regenerate = zero LLM calls), deterministic fallback; difficulty anchors ship to the grader prompt (a-5 same bar, testable)
    • Oral defense (REQ-3-006): seventh Examiner agent grounded in digest + variant statement (learner-anonymous prompts), mock-first VoiceProvider + browser-native fallback (CUT-1: real server STT/TTS is a v0.4 seam), DefenseStore transcript + A-109 long-pause integrity signals, per-turn latency instrumentation
    • Learner surface integration (REQ-3-007/008): /build is a real environment (variant + starter files in a live sandbox, file CRUD, Run/Test with read-only output per CUT-2, live telemetry pulse, Lab on the live trace); /defend is a live defense + Grade My Work renders real rubric bars. E2E credential flow proven over real uvicorn + real namespaces

    Honest limitations (G-6)

    (a) Lab/Assessor/Proctor run on real engine inputs — the v0.2 mock caveat is retired. (b) Sandbox scope = coding IDE only; design tool + simulation are v0.4 (D-025). (c) Identity/age-gating (KYC) deferred per founder directive — age-gating remains the v0.1 visual flow; abuse control ships instead (G-5). (d) Voice is mock-first with browser-native fallback; real server STT/TTS deferred to v0.4 (CUT-1). (e) Sandbox resource limits are PARTIALLY enforced: memory/CPU/single-file/wall-clock kernel-enforced via rlimits; per-sandbox pids + hard disk quota are NOT (no cgroup delegation on this box) — mitigated by the 512MB workdir sweep + per-learner caps. (f) Exec results reach the trace as capture-agent-observed file_diff/activity; command/run/test events fire only in agent REPL mode (grading digests under-represent process on the exec path) — a seq-lease protocol lands in v0.4. (g) Starter tests cannot pass on this box (no pytest in the exec namespace's Python — environmental). (h) Userns is not a DAC write barrier — a per-sandbox runtime uid is the v0.4 hardening.

    Quality gates

    407 tests green · pnpm build 4/4 · typecheck 7/7 · ruff clean · P7 review fixed 8 defects (CORS PUT preflight, WS origin gate, symlink escape, ingest session leak, O(n²) flood count, retry sandbox leak, flake de-rat) with 10 regression tests · audit re-pointed two orphaned tags and cleared doc drift across 7 files

    Phases: v0.2.1 pre-execution · v0.2.2 sandbox fabric · v0.2.3 telemetry · v0.2.4 grading · v0.2.5 variants · v0.2.6 voice defense · v0.2.7 integration · v0.2.8 milestone release

    Downloads