Roadmap

Status: pre-1.0, v0.0.0 (branch upgrade/top-1pct). The product-roadmap phases below (0–5) are code-complete in the monorepo — scaffold, MVP spine, all surface + connector adapters, the BENCHMARKS gate, and streaming/provider breadth (surfaces take an injected transport; production wiring is per-operator). On top of that, the canonical ChatModel contract and the @tenet/agent orchestrator (runAgent drives the fail-closed turn end-to-end) are shipped. The current-state banner just below is the single canonical live status. Dates are targets, not promises.

Source of detail for each row: ARCHITECTURE.md (component map) and BENCHMARKS.md (measurable targets).

State as of 2026-07-18 (branch upgrade/top-1pct): 1328 tests passing across 98 suites, all green; 63 workspace packages. CI green. The 5 deferred review findings from the adversarial pass on c9c9fb0 all shipped. The 7 open research questions all closed in RESEARCH-PASS-2.md.

Canonical-contract upgrade (complete): @tenet/core now defines ONE ChatModel contract — a ModelMessage[] array, a REQUIRED AbortSignal, tool round-trips (ContentBlock: text tool_use tool_result), and a lossless StopReason union (packages/core/src/model.ts). It replaced four duplicated single-string chat({system, user}): Promise<string> interfaces. All 6 model adapters (anthropic, bedrock, google, mistral, ollama, openai) migrated, fixing three provider “abnormal stop” bugs that had reported an abnormal stop as a clean completion: Anthropic streaming truncation, Mistral model_length context overflow (models/mistral/src/index.ts:60), and Google content-block reasons — PROHIBITED_CONTENT / SPII / BLOCKLIST / IMAGE_SAFETY (models/google/src/index.ts:73-76) — now map to max_tokens / refusal instead of end_turn. @tenet/verifier, @tenet/router, @tenet/streaming (re-exports the canonical StreamChunk), @tenet/ag-ui, apps/quickstart, apps/measure-real, and apps/community-bot all speak the canonical contract; the transitional migration bridges were deleted — every consumer speaks the canonical contract, no bridges remain. The ${role}: ${content} turn-spoof is gone from community-bot — history is structured ModelMessage[], so a user message containing assistant: can no longer forge a prior assistant turn (apps/community-bot/src/index.ts:114), guarded by a wire-level role-spoof regression test in the Anthropic adapter.

Agent orchestrator (shipped): the @tenet/agent package drives the full turn. runAgent (packages/agent/src/orchestrator.ts) executes OrchestratorState as a @tenet/workflow graph — injectionGate → retrieveNode → cacheNode → compactNode → criticLoop, all fail-closed (compactNode is an opt-in long-history compaction pre-step — a compaction error abstains rather than send a truncated prompt). The decides-and-drafts Reasoner (modelReasoner + parseEnvelope) is grounded-or-abstain (no answer without a non-blank citation; every ambiguity → abstain). verifyAndEmit is the single emit edge (verify → ≥1 grounded citation → outbound leak gate); tools run through governance (runTools); terminals are carried in state.halt with a finalize default + a top-level catch that fails closed on any error. A failed draft is rewritten (deps.rewrite) and re-verified through the same emit edge up to maxRepairRounds, then abstains. Follow-up (does not weaken the guarantee): real-model validation of the decides-and-drafts prompt.

Phase 0 — Scaffold (DONE)

  Done
Repo layout, license, security policy ✓
pnpm workspaces, TS strict + ESM ✓
CI: typecheck + test + audit + Trivy + eval-regression (lint is a no lint yet no-op stub — real ESLint is Phase 6.1) ✓
@tenet/core types (Conversation, Surface, Filter, Secret, PathHandle, OutcomeEvent) ✓
@tenet/verifier atomic-claim multi-judge ✓
@tenet/policy Constitutional Principles registry + composer + drift fingerprint ✓
@tenet/telemetry OTel GenAI semconv + outcome events + CostMeter ✓
Adversarial code review pass + 6 confirmed bugs fixed ✓
118 unit tests across 8 suites ✓
docs/ARCHITECTURE.md + docs/BENCHMARKS.md ✓
CONTRIBUTING / CODE_OF_CONDUCT / issue + PR templates ✓

Phase 1 — MVP (LANDED — code complete; production wiring TBD)

The smallest end-to-end vertical that proves the spine. From ARCHITECTURE.md §12.

  Status
@tenet/retrieval — Bm25Index + InMemoryVectorStore + RRF fusion + HybridPipeline ✓
@tenet/rate-limit — TokenBucket + CircuitBreaker + RateLimitScheduler + 5 platform policies ✓
@tenet/guardrails — PII redactor (7 kinds), composable filters ✓
@tenet/router — AdaptiveRouter + InMemorySemanticAnswerCache + SpeculativeAgent ✓
@tenet/memory — WorkingMemory + InMemoryEpisodicMemory + InMemorySemanticMemory ✓
surfaces/telegram — TelegramSurface adapter (HTML formatter, citation rendering, rate-limit gate) ✓
models/bedrock — AnthropicOnBedrockChatModel (no AWS SDK hard dep) ✓
stores/vector/pgvector — PgVectorStore adapter (no pg SDK hard dep, all values parameter-bound) ✓
stores/state/redis — InMemoryStateStore + RedisStateStore adapter ✓
apps/community-bot — CommunityBot reference composition ✓
eval/harness — runEval + builtinAssertions + regressionGate ✓
Golden eval dataset (~100 cases) committed not started — Phase 2
End-to-end smoke run against real Bedrock not started — Phase 2

MVP success criteria (from ARCHITECTURE.md §12):

  1. Conversations replay through the new runtime end-to-end. → composition shipped in @tenet/app-community-bot; real-LLM smoke run pending.
  2. CoVe + Reflexion + citation enforcement run on every output (not opt-in). → atomic-claim multi-judge verifier wired; CoVe / Reflexion as JudgePromptDecorator strategies (B5).
  3. Telegram rate-limit policy enforced — simulate flood, observe queueing. → RateLimitScheduler with TELEGRAM_POLICY; integration test pending.
  4. OTel spans visible in a local Jaeger. → semconv keys defined in @tenet/telemetry; exporter wiring pending.
  5. CI runs eval suite as a gate. → @tenet/eval-harness + regressionGate ship the primitive; CI workflow integration pending.

Phase 2 — Surfaces + adapters expansion (SHIPPED except where noted)

AS-BUILT status, verified against the monorepo 2026-07-18. ✓ = adapter/code present; surfaces take an INJECTED transport (no hard SDK dependency), wired to the concrete client by the operator at composition. The current-state banner at the top of this doc is the single canonical live status.

  Status
surfaces/discord (injected client, no discord.js hard dep) ✓
surfaces/slack ✓
surfaces/web-widget (SSE + JWT) ✓
models/anthropic — canonical ChatModel + chatStream (SSE), StopReason mapping ✓
stores/vector/qdrant ✓
@tenet/memory-adapters — mem0 / Letta / Zep providers ✓
Eval set growth to 1k+ cases via assertion mining not started

Phase 3 — Enterprise (SHIPPED except where noted)

  Status
surfaces/teams (Bot Framework + Adaptive Cards) ✓
surfaces/rest (OpenAPI 3.1) ✓
surfaces/grpc ✓
Ticketing webhook receiver (connectors/ticketing, HMAC-verified) ✓
connectors/ticketing — zendesk / intercom / freshdesk / servicenow modules ✓
apps/enterprise-support — full multi-surface reference ✓
On-prem Helm chart (infra/helm/tenet) ✓
SOC2 audit prep not started (process; no code artifact)
MCP tool gateway with OAuth (@tenet/tools-mcp-gateway) ✓
WASM-sandboxed tool execution (@tenet/tools-wasm-sandbox + -wasmtime) ✓

Phase 4 — Beat the benchmarks (DONE)

Every row in BENCHMARKS.md measured + CI-gated:

Phase 5 — Streaming + provider breadth (DONE)

Phase 6 — Pre-tool-call governance + memory consistency (DONE)

Two first-class layers most agent frameworks leave to custom integration — evaluating a tool call against policy before it executes, and read-after-write consistency for agent memory — ship here as code:

Phases 7–14 — Breadth + protocols + security (DONE)

Shipped 2026-06-04: Gemini/Mistral/Ollama adapters, Bedrock + Anthropic streaming, OTLP wiring, voice surface (@tenet/voice + OpenAI Realtime + Twilio Media Streams), SKILL.md loader, Matrix surface, Dockerfile + health endpoints + pre-commit, @tenet/tool-call-repair + @tenet/net-policy, @tenet/acp (Agent Client Protocol), and a 23-finding vuln-test triage (44 bugs caught cumulatively across three adversarial review passes).

Phase 15 — Top-0.01% market gaps (DONE, 2026-06-10)

Fresh A-to-Z audit (3 parallel reviewers) + live June-2026 market research identified the three TypeScript open lanes; all shipped:

Deferred from adversarial review — ALL LANDED

Surfaced by the 2026-06-04 review pass, all five shipped:

Anti-goals (will NOT do)