Splunk Observability Cloud OTLP exporter — wraps @crewhaus/otel-exporter with X-SF-TOKEN auth + realm routing (Section 37)

Cross-harness discovery + fleet inventory/health rollup — the library core behind crewhaus fleet (discoverHarnesses, buildFleetInventory, buildHarnessHealth, describeFleetExit, seam-injected bulk runner)

NLG-metric graders: ROUGE-1/2/L, BLEU-1..4, METEOR. Pure-TS; pluggable into @crewhaus/grader-registry (Section 38)

Envelope encryption for hash-chained audit records: per-tenant DEK wrapped by KEK from @crewhaus/secrets-manager (Section 39)

Pillar-2 orchestration — wires eval-runner (fitness) + prompt-optimizer (search) + spec-patch (mutation) into a single 'crewhaus optimize' workflow that closes the active-optimization loop.

Track H / §59 — emits an Anthropic-compatible Claude Code plugin directory from any CrewHaus IR variant. Source: claude-plugins-official (Anthropic).

bun build --compile wrapper producing self-contained crewhaus binaries for linux/macos/windows × x64/arm64; auto-generated Homebrew/apt/Scoop/Winget manifests (Section 32)

Browser-based REPL for the CrewHaus meta-harness compiler — Monaco editor + live trace timeline + OAuth quota + cross-tenant isolation (Section 35).

Schema-driven form + loop-projection model for CrewHaus spec editing — the shared authoring engine behind the iPad Studio and the local Studio (Section 26 Studio)

Spec authoring + inline trace inspection + run-from-editor for the CrewHaus meta-harness compiler (Section 35).

Drop-in web UIs for every CrewHaus harness shape, plus a one-command scaffolder. Compile a spec, point @crewhaus/ui at it, get a polished shape-aware UI in the browser.

v0.3.0 Goal 5 (design §6): scheduled memory consolidation — a deterministic phase-1 pass (TTL sweep, dedupe/supersede, staleness flags, sessions-index fold-in, proof pinning, handoff refresh, trash purge) plus ONE bounded, budget-capped model synthesis se

0.3.0 memory release §3.1 — local update-in-place wiki: versioned markdown articles (supersede, never delete), optimistic concurrency (stale_article_version), hybrid BM25+embedding recall with one-hop [[wikilink]] expansion.

Durable per-(route-key, model) reward scoreboard that makes model_pool routing improve with usage (Section 17)

v0.3.0 composition root (design §1 principle 1): wireMemory/wireContinuity/wireWiki construct the memory fabric's stores, register its tools, and return runChatLoop-ready seams — one stable call replacing per-emitter memory codegen.

Resolve agent.model strings to a lazy-loaded ProviderAdapter (Section 17)

Shipped default skills (continuity, learning-loop, dream) and builtin slash commands

IntelliJ / WebStorm / PyCharm parity for the CrewHaus VS Code extension: spec autocomplete + lint + Run Spec / Run Eval / Run Canary configurations + spec-registry tool window (Section 35).

Platform-neutral agent-loop runtime (tools + budget + limits + trace) with all host services injected — imports no node:* builtins, runs on Cloudflare Workers

Human-rating distillation: FeedbackRecords + session turns to eval samples, the auto-distill watermark, and the daemon janitor step that keeps a <spec>-ratings registry dataset current