-
coreci-bot released this
2026-09-12 02:48:54 +00:00 | 36 commits to main since this releaseMilestone v0.3 Credential Engines — Phase 3 Process-Trace Grading Engine shipped.
- Deterministic trace digest (D-028): test pass/fail, edit count, error/fix cycles + fix latency, idle gaps, command categories — computed in code; the LLM sees ONLY the compact digest (leak-tested with planted markers through the full stack)
- Rubric scoring through the reused D-020 4-layer JSON defense; validated RubricScore (4 criteria x 0-4 + strengths/gaps/verdict)
- G-4 gate (binding): gapped or INCOMPLETE_FLOODED traces → first-class
UNGRADABLE_TRACE_INCOMPLETE, LLM never called — no credential from an incomplete trace - Calibration contract: strong >= lazy on process quality, strong > struggling on correctness — deterministic digest-feature separation over D-021-aligned fixtures
- GradeStore (SQLite, protocol-wrapped, D-027) + POST/GET /v1/assessment/grade
283 tests green; ruff clean; verified end-to-end (real WS ingest → HTTP grade).
Downloads
-
v0.2.3 — v0.3 Phase 2 (live build telemetry) Pre-Release
2026-09-12 01:41:17 +00:00 | 44 commits to main since this releaseMilestone v0.3 Credential Engines — Phase 2 Live Build Telemetry shipped.
- First real persistence (SQLite/SQLModel, WAL) behind a Postgres-migration-ready
TraceStoreprotocol (D-027) - WebSocket ingest
/v1/telemetry/ingestwith server-side(learner,task,seq)dedup, gap detection, keepalive (D-026) - Flood control (G-3): >50k events/task → WS 1008 + trace flagged
INCOMPLETE_FLOODED(terminal, grader refuses); silent drops forbidden - Stdlib-only in-sandbox capture agent: fsync spool, exponential-backoff reconnect, ordered flush; lifecycle tied to sandbox
- Telemetry-wired sandboxes:
create(learner_id, task_id)spawns the agent in a persistent helper/inner namespace topology (offline inner ns; agent joins mount ns only, stays online) - Durability proven end-to-end: real uvicorn + real namespace + severed connection → every event stored exactly once, in order
- TS mirror types for the web surfaces (packages/types/telemetry.ts)
218 tests green; ruff clean. Verifier caught + fixed a namespace-init orphan leak in destroy() (with regression test). Known P1 for Phase 3: live traces carry activity/file_diff kinds; command/stdout kinds flow through the agent REPL.
Downloads
- First real persistence (SQLite/SQLModel, WAL) behind a Postgres-migration-ready
-
v0.2.2 — v0.3 Phase 1 (sandbox fabric) Pre-Release
2026-09-12 00:47:45 +00:00 | 53 commits to main since this releaseMilestone v0.3 Credential Engines — Phase 1 Sandbox Fabric shipped.
SandboxBackendprotocol +UnshareBackend(Linux user/mount/pid/net namespaces) — probe-verified on this box (in-ns uid=0, network isolated, writes contained)- Lifecycle manager + concurrency guard (503 when full) + startup orphan reaper
- Resource limits: memory/CPU/single-file rlimits + wall-clock reaper kernel-enforced; 512MB workdir-size sweep; per-sandbox pids/hard-disk-quota documented as accepted v0.3 gaps (no cgroup/sudo on box)
- Abuse control (no-auth): learner allowlist (403), per-learner cap (429), create-rate cap (429)
174 tests green; ruff clean;
/v1/sandboxeslifecycle verified live. Known v0.3 gap: userns is not a DAC barrier + killing the supervisor doesn't kill in-ns children — disclosed for full hardening in the post-MVP containerd backend.Downloads
-
v0.2.1 — v0.3 Phase 0 (pre-execution) Pre-Release
2026-09-12 00:47:45 +00:00 | 60 commits to main since this releaseMilestone v0.3 Credential Engines — Phase 0 pre-execution complete.
- SPECIFY: REQ-F-007/008/009/010/021 activated as REQ-3-001..008; REQ-F-017 (KYC/age-gating) deferred per founder directive
- CLARIFY: 12 ambiguities auto-resolved (A-101..112)
- RESEARCH: sandbox isolation probe-verified via unshare Linux namespaces (no container runtime/sudo on box); D-024..032
- PLAN: 6 phases / 20 waves / 39 tasks
- GRILL: GO-WITH-CHANGES — binding decisions G-1..G-6 + scope cuts CUT-1 (real STT/TTS → v0.4) / CUT-2 (interactive shell relay → v0.4) applied
- MVP/UX gate: passed
Phase 0 patch on the v0.2 line. Execution phases follow on milestone/v0.3-credential-engines.
Downloads
-
2026-09-11 17:35:00 +00:00 | 71 commits to main since this releaseNextcraft v0.2.0 — AI Tutor Architecture
The six AI tutor agents are live: a Python FastAPI service (
apps/ai-service) hosting Coach, Tutor, Lab, Assessor, Proctor, and Mentor behind a provider-agnostic LLM layer, wired into the v0.1 learner surface with real streaming.Honesty note (per D-015 and G-5)
Lab, Assessor, and Proctor operate on mock engine inputs — simulated telemetry timelines and pre-baked artifacts/transcripts. Their real engines (sandbox fabric, process-trace assessment, identity verification) are v0.3+. All six agents make real LLM calls; only the inputs to three of them are mocked.
The Six Agents
- Coach — pacing, motivation, retrieval practice; ends every reply with exactly one next action
- Tutor — one concept per reply, worked example, Socratic check
- Lab — in-flow feedback on build-session telemetry (mock inputs)
- Assessor — rubric grading of artifacts + defense transcripts with structured, evidence-cited scores (mock inputs)
- Proctor — integrity signals with coaching-shaped interventions; never punitive (mock inputs)
- Mentor — long-horizon career narrative, session-backed
Architecture
- Provider-agnostic LLM layer — OpenAI-compatible httpx client: ollama-cloud (default, gemma4:31b), local endpoint, deterministic mock (tests/CI — the suite never calls the cloud)
- D-016 SSE envelope — meta → deltas → done →
[DONE], in-band errors, ping keep-alive immunity - 4-layer structured output defense — response_format auto-degrade → prompt-embedded schema → defensive parse → bounded retry
- Session store — agent-scoped, 20-message windows, LRU cap, DB-migration-ready
Learner Surface (all streaming, all with error+retry states)
- Dashboard: Coach ↔ Tutor switcher chat + Mentor career panel
- Byte viewer: Tutor concept walkthroughs · Sandbox: Lab feedback panel · Defense: Assessor rubric bars + Proctor integrity banner
Phases (all shipped)
P0 pre-execution → v0.1.1 · P1 scaffolding → v0.1.2 · P2 agent framework → v0.1.3 · P3 Coach+Tutor → v0.1.4 · P4 Lab+Assessor → v0.1.5 · P5 Proctor+Mentor → v0.1.6 · P6 learner surface → v0.1.7 · P7 final review → v0.2.0
Quality
134/134 tests green (mock provider only) · ruff clean · tsc green across workspaces · next build 15/15 pages · E2E smoke 7/7 · final adversarial review: GO (after CRLF wire-format and disconnect-handling P0 fixes, regression-covered) · project health audit: PASS 7/7 · secrets never in code/logs/commits/history
Escalation record
P1 executor delegation failed twice (empty subagent results); auto-resolved at full autonomy to inline execution with identical plan fidelity (commit 3271373, reflog-only after squash-delete — recorded here durably per audit remediation).
Requirements
All 12 v0.2 requirements complete (REQ-2-001..012). 40 requirements complete across v0.1+v0.2.
Next (roadmap candidates, v0.3+)
Assessment engine (process-trace grading, per-learner variants) · sandbox fabric · identity verification & age-gating · payments
Downloads
-
2026-09-11 16:36:21 +00:00 | 64 commits to main since this releaseNextcraft v0.1.7 — v0.2 Phase 6
The v0.1 prototype is now alive: all four learner surfaces stream from the real AI service.
Delivered (REQ-2-011, REQ-2-012)
- useChatStream hook — fetch + ReadableStream SSE client: byte buffering with frame reassembly across chunk boundaries,
: pingkeep-alive immunity, multi-line data joining, idempotent AbortController cleanup (Strict-Mode safe),[DONE]sentinel handling - Dashboard chat — real token-by-token streaming with a Coach ↔ Tutor switcher (agent-scoped sessions, no persona bleed), streaming cursor, error state with retry when the service is down
- Byte viewer — Tutor panel: Socratic concept walkthroughs per competency
- Sandbox — Lab in-flow feedback panel on the telemetry sidebar
- Defense — Assessor rubric output as live score bars + strengths/gaps, and a Proctor integrity banner with severity chips and a coaching-shaped suggestion
- Dashboard Mentor panel — long-horizon career narrative stream
- A11y + dark mode throughout (tablist/aria-selected, role=alert, progressbar semantics, full dark: coverage)
Quality: 131/131 ai-service tests, tsc green across workspaces, next build 15/15 pages, E2E smoke 7/7 over live HTTP (chat routing, lab, mentor, assessor/proctor paths, CORS).
Next
Final phase: multi-persona code review, project health audit, and the v0.2.0 milestone release.
Downloads
- useChatStream hook — fetch + ReadableStream SSE client: byte buffering with frame reassembly across chunk boundaries,
-
2026-09-11 16:20:57 +00:00 | 65 commits to main since this releaseNextcraft v0.1.6 — v0.2 Phase 5
The six-agent roster is complete.
Delivered (REQ-2-009, REQ-2-010)
- Proctor agent (persona proctor-v2) — classifies integrity signals from defense-session telemetry (tab switches, idle gaps, large pastes, focus loss) into typed/severity/noted signals; recommends exactly ONE supportive coaching intervention. Never punitive, assumes good faith. Mock telemetry inputs — real identity/attention signals are v0.3+ (D-015)
- Mentor agent (persona mentor-v2) — long-horizon career narrative: where the learner is (evidence-based), what their stack unlocks next, how they position in the AI-era labor market. Streaming, session-backed with follow-up context
- Endpoints —
POST /v1/proctor/signals(validated JSON, 404/422/502 paths);POST /v1/mentor/narrative(SSE, D-016 envelope, session persistence) - Registry complete — all six agents resolve centrally: coach, tutor, lab, assessor, proctor, mentor
Quality: 131/131 tests green (cloud-free), ruff clean, live-server smoke verified both endpoints.
Next
Phase 6: learner surface integration — wire the v0.1 UI to the real streaming service.
Downloads
-
2026-09-11 16:16:51 +00:00 | 66 commits to main since this releaseNextcraft v0.1.5 — v0.2 Phase 4
The build-and-grade agents join the roster — real LLM logic over mock engine inputs (simulated telemetry, pre-baked artifacts; real engines are v0.3+, per D-015).
Delivered (REQ-2-007, REQ-2-008)
- Lab agent (persona lab-v2) — streams concrete in-flow feedback from sandbox telemetry: what the events show, the one most useful adjustment, one next step. Coaching tone on suspicious patterns (huge paste → self-check suggestion, not accusation)
- Assessor agent (persona assessor-v2) — applies rubrics to artifacts + defense transcripts; returns pydantic-validated structured scores with cited evidence, strengths, gaps, verdict; the 4-layer JSON defense (response_format degrade → prompt schema → defensive parse → bounded retry) is fully exercised
- Endpoints —
POST /v1/lab/feedback(SSE, D-016 envelope, 404 unknown scenario);POST /v1/assessment/evaluate(validated JSON, 404/422/502 paths) - Mock input corpora — 3 telemetry scenarios (strong/struggling/flagged), 2 pre-baked artifacts with defense transcripts, 2 rubrics; IDs string-identical on the TS and Python sides (D-021, mechanically cross-checked)
Quality: 110/110 tests green (cloud-free), ruff clean, tsc green, lockfile integrity restored (apps/ai-service importer).
Next
Phase 5: Proctor + Mentor agents — integrity signals with coaching interventions, long-horizon career narrative.
Downloads
-
2026-09-11 15:59:57 +00:00 | 67 commits to main since this releaseNextcraft v0.1.4 — v0.2 Phase 3
The first two fully-implemented AI tutor agents, live and streaming.
Delivered (REQ-2-005, REQ-2-006)
- Coach agent — pacing guidance, motivation tied to trajectory, retrieval-practice questions; always ends with exactly one concrete next action (persona coach-v2)
- Tutor agent — one concept per reply, worked example first, Socratic check question before advancing (persona tutor-v2)
- Agent routing —
/v1/chat/streamroutes by agent name; unknown agents rejected 422; meta event carries the routed agent; sessions scoped per agent (no persona bleed) - Live-verified personas — manual ollama-cloud probe (gemma4:31b): Coach responded with pacing + retrieval question + one action; Tutor with concept + worked example + Socratic question. Distinct voices, on-persona.
Quality: 80/80 tests green (mock provider only — suite never calls the cloud), ruff clean, boundaries verified.
Next
Phase 4: Lab + Assessor agents over mock engine inputs (simulated telemetry, pre-baked artifacts).
Downloads
-
2026-09-11 15:52:00 +00:00 | 68 commits to main since this releaseNextcraft v0.1.3 — v0.2 Phase 2
The shared framework all six AI tutor agents build on.
Delivered (REQ-2-004)
- BaseAgent ABC — system_prompt / build_messages / stream_reply / structured_reply; default provider delegation
- Session store — protocol + in-memory implementation: asyncio.Lock-guarded, 20-message rolling window, 500-cap LRU eviction, agent-scoped sessions (DB-migration-ready)
- Agent registry — central registration, duplicate/unknown error paths
- Structured output defense — 4 layers: response_format with auto-degrade, prompt-embedded schema, fence-strip/first-balanced-object parse, single bounded retry
- Prompt library — six versioned modules with learner-context renderers
- Learner context corpus — pydantic-typed, IDs string-identical to the TS mock-data layer
- Chat endpoint session wiring — persisted turn history, windowed replay, assistant-turn persistence
Quality: 61/61 tests green, ruff clean, module boundaries verified (agents/ never imports api/).
Next
Phase 3: Coach + Tutor agents — full LLM implementations with agent routing.
Downloads