Pending — Research + Architecture Triage

As of: 2026-07-18 (revised after the canonical-ChatModel upgrade on upgrade/top-1pct). Lists what’s queued, why, and how I’d rank it.

Upgrade update (2026-07-18). Phase 1 (canonical ChatModel migration) landed and Phase 2’s agent turn is SHIPPED. B2 and B3 are now RESOLVED (see bucket B); the single-string ChatModel schism is resolved by one messages-array contract (packages/core/src/model.ts); the model-adapter C-items are done (all 6 adapters exist). The @tenet/agent package now ships the full driven turn: runAgent (packages/agent/src/orchestrator.ts) executes OrchestratorState as a fail-closed @tenet/workflow graph (nodes → Reasoner → governance-gated tools → verify → single emit edge). B1, B4, B5 remain open. Details inline below.

All 26 original rows closed. See RESEARCH-PASS-2.md for buckets A + D (research), CHANGELOG.md for bucket B (5 adversarial fixes) + bucket C (11 MVP packages). This doc is preserved as the historical triage; future open items live in GitHub Issues.

Companion to ROADMAP.md. This doc is the “what to read / build / verify next.” Roadmap is the “what ships in which phase.”

Bucket A — Open research questions (carried from pass 1)

Pass 1 of the deep-research workflow surfaced these as unverified after 3-vote adversarial verification. They affect specific phase-2/3 architecture decisions; we move forward without them at our own risk.

#	Question	Why it matters	Cost to research	Suggested when
A1	Closed-tool architecture clues (Decagon / Sierra / Cresta / Aisera / Moveworks): model routing, retrieval stack, eval harness, multi-tenancy, hand-off model.	Decides whether `@tenet/router` ships fine-tune-aware routing by phase 2 or defers.	medium (engineering blogs, job posts, conference talks, patents)	before phase 2
A2	Vector store cost-perf at 10M-1B vectors: pgvector vs Qdrant vs Weaviate vs Milvus vs LanceDB vs Pinecone vs Turbopuffer vs Vespa.	Decides phase-2 default. pgvector is the current default by ease-of-self-host, not by measured perf.	high (need own benchmarks or reproducible third-party)	before phase 2 vector-store default lock-in
A3	MCP enterprise adoption signal in named Fortune-500.	Decides whether MCP-native is a launch differentiator or a phase-2 story.	low (announcements, GH issues, vendor pages)	before v1.0 marketing
A4	Eval / observability layer head-to-head: LangSmith vs Braintrust vs Arize Phoenix vs Langfuse vs Helicone + OTel GenAI conformance.	Decides our default recommendation for operators. OTel export covers all of them, but docs need to pick one.	low-medium	before phase-1 eval-as-CI

Bucket B — Adversarial review findings deferred from f88a343

From the 6-agent adversarial pass on the c9c9fb0 implementation. Fix #1, #2, #3, #6, #7, #10 landed. B2 and B3 were resolved by the 2026-07-18 upgrade; B1, B4, B5 remain.

#	Status	Finding	File	Severity	Cost to fix
B1	OPEN	`buildCitations` 30-char substring head-match drops citations for paraphrased claims.	`packages/verifier/src/verifier.ts:127`	medium	small — replace with swappable `pickBackingSource(claim, sources)` strategy in `@tenet/core`
B2	RESOLVED	`PathHandle` brand type now has a sanctioned `openPath()` factory (allow-list root via `configurePathRoot()`, absolute-path + `..`-traversal + root-escape rejection). No longer vapor type safety.	fixed in `packages/core/src/path.ts:58` (`openPath`); brand at `packages/core/src/types.ts:34`	—	done
B3	RESOLVED	The canonical `ChatModel` now takes a required `AbortSignal` (`ChatRequest.signal`, non-optional) — its presence is enforced by the type system, so a cancelled turn can release the in-flight call.	fixed in `packages/core/src/model.ts:57` (`readonly signal: AbortSignal`)	—	done
B4	OPEN	`OutcomeEmitter` sink-isolation `try/catch` swallows infinite recursion + all programmer mistakes silently.	`packages/telemetry/src/index.ts:76`	medium (operational opacity)	small — add re-entrancy guard + log swallowed errors at warn (without recursing back through emit)
B5	OPEN	`VerifierConfig` bakes app-specific concepts (`allowedUrlPatterns`, `adversarialExamples`) into framework.	`packages/verifier/src/types.ts:28`	medium (architecture / altitude)	medium — replace with pluggable `ClaimPreFilter` + `JudgePromptDecorator` interfaces

Bucket C — MVP packages not yet implemented

Each is in ROADMAP.md Phase 1. All 11 packages landed (see CHANGELOG.md); the model-adapter and community-bot rows below were re-verified against the code by the 2026-07-18 upgrade.

#	Status	Package	Why it’s blocking MVP	Estimated scope
C1	landed	`@tenet/retrieval`	RAG is the heart of grounded answers — no retrieval, no source-grounding to enforce.	large (chunker, contextualizer, BM25 + dense, RRF fusion, rerank, swappable vector-store adapter)
C2	landed	`@tenet/rate-limit`	Without it any surface adapter has to roll its own; first PR breaks the architecture.	medium (token bucket per (platform, tenant, channel, user), circuit breaker on 401/403 streak)
C3	landed	`@tenet/router`	Adaptive cost-aware routing is core to the cost-per-resolution target.	medium-large (classifier, semantic answer cache, speculative streaming)
C4	landed	`@tenet/memory`	Conversation history can use a stub initially; long-term episodic + semantic is phase 2.	medium (episodic + semantic + working interface; adapters can land later)
C5	landed	`@tenet/guardrails`	PII redaction + input/output filters. Required before any prod traffic.	medium
C6	landed	`surfaces/telegram`	First reference surface. Proves the adapter pattern.	medium (grammY adapter, formatter, rate-limit policy wiring)
C7	DONE	`models/bedrock`	First model adapter. Implements `ChatModel`.	done — all 6 adapters exist (`models/{anthropic,bedrock,google,mistral,ollama,openai}/src/index.ts`), migrated to the canonical `ChatModel` contract.
C8	landed	`stores/vector/pgvector`	First vector store. Implements the swappable interface.	medium
C9	landed	`stores/state/redis`	Session + cache state.	small
C10	DONE	`apps/community-bot`	First reference deployment.	`apps/community-bot/src/index.ts` exists and is on the canonical contract; history is now STRUCTURED `ModelMessage`s (`textMessage(m.role, m.content)` at `index.ts:114`), killing the `${role}: ${content}` turn-spoof. End-to-end orchestrator now shipped (`runAgent`, see below).
C11	landed	`eval/harness`	Eval-as-CI requires a runner before any eval-set growth makes sense.	medium-large (golden dataset loader, LLM-as-judge, regression gate)

Upgrade status — canonical `ChatModel` (Phase 1) + fail-closed Reasoner (Phase 2)

(2026-07-18, branch upgrade/top-1pct.)

Landed:

Phase 1 — one canonical ChatModel contract in @tenet/core. A messages array (ModelMessage[]), a required AbortSignal, tool round-trips (ContentBlock: text tool_use tool_result), and a lossless StopReason union (packages/core/src/model.ts). This replaced the duplicated single-string chat({system, user}): Promise<string> interfaces — the schism the recon flagged as the root of “no conversation memory.” Role structure now reaches the model. All 6 adapters were migrated onto it, and the transitional bridges were deleted (no fromLegacyModel/asLegacyModel/LegacySingleStringModel/LegacyChatArgs remain in code — every consumer speaks the canonical contract).

Phase 2 — the @tenet/agent driven turn (SHIPPED). runAgent (packages/agent/src/orchestrator.ts) executes OrchestratorState as a @tenet/workflow graph: injectionGate → retrieveNode → cacheNode → compactNode → criticLoop, all fail-closed (compactNode = opt-in long-history compaction; a compaction error abstains, never sends a truncated prompt). The FAIL-CLOSED 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.

Still pending (stated honestly):

Real-model validation of the decides-and-drafts prompt (buildSystem) — deferred until an API key exists in the build env. The control-flow guarantee does not depend on it. (The answer repair-retry — RewriteFn + re-verify in the criticLoop answer branch — is now shipped.)
The verifier’s own fail-CLOSED MODE ships (Phase 3.1 — VerifierConfig.failClosed: extractor error / judge error / judge-garbage / a truncated (max_tokens) or over-cap extraction → not-supported). The truncation + maxClaims-drop hardening (3.3) now ships too — verified in the tree: claimExtractor.ts throws under failClosed on both an over-cap AND a max_tokens-truncated extraction, and reasoner.ts abstains on a truncated draft. Still ahead: the non-numeric claim tier (3.2 — named-entity coverage + negation/scope guard).
REAL-MODEL benchmark numbers still do not exist (no model access in this environment). The hermetic stub remains the only measured plane; keep the stub-vs-real split in the benchmark docs honest — no real-model numbers are claimed.

Bucket D — Fresh research worth doing on current state of the field

Pass 1 was on 2026-06-03. The agent-framework field churns monthly. These are worth a focused pass before phase 1 implementation locks in choices.

#	Topic	Reason	Estimated cost
D1	Anthropic Claude Agent SDK + Opus 4.8 + Haiku 4.5 — actual SDK feature surface as of today (compaction, batch, structured outputs, prompt caching, MCP integration).	We picked Opus 4.8 + Haiku 4.5 in BENCHMARKS.md but didn’t verify the SDK shape. Affects `models/anthropic` and the router.	low
D2	MCP spec changelog since pass 1 — transport stability (stdio vs SSE vs Streamable HTTP), OAuth ratification status, server-registry consolidation.	“MCP-native” is a marketed differentiator; spec drift would change the story.	low
D3	OpenTelemetry GenAI semconv stable version + which exporters fully implement it.	We claim “OTel GenAI semconv native” — need to confirm what version we conform to.	low
D4	New / evolved OSS frameworks: Letta long-term-memory, Mastra TS-first, Pydantic-AI v1, Inkeep, Smolagents, OpenAI Agents SDK.	We need to know what to compete on.	medium
D5	Vectara HHEM-2.1 vs newer hallucination judges (latest leaderboard).	Hallucination CI gate depends on this choice.	low
D6	Cohere Rerank 3.5 vs Voyage Rerank vs BGE-Reranker-v2-M3 vs jina-reranker-v2.5 (head-to-head).	Reranker is part of the contextual-retrieval recipe; current pick was BENCHMARKS.md-era.	low

Suggested execution order

Strict-precedence rule: if a row in a later bucket can fail because of an unresolved row in an earlier bucket, do the earlier one first.

D3, D5, A4 (~15 min research total) — pick OTel semconv version, hallucination judge, eval-layer recommendation. These unblock Bucket C without code.
~~B3 (AbortSignal)~~ — DONE (packages/core/src/model.ts:57, required signal).
~~B2 (PathHandle factory)~~ — DONE (packages/core/src/path.ts:58, openPath).
C7 + C8 + C2 (models/bedrock + pgvector + rate-limit) — DONE; all three packages landed.
~~C1 (@tenet/retrieval)~~ — DONE; package landed.
D4 (refreshed OSS framework landscape) — feeds the next architecture review.
A2 (vector-store benchmark) — only matters once we have non-trivial corpus. Defer until MVP shows real traffic.
A1 (closed-tool architecture) + D1, D2, D6 — informs phase-2 polish, not phase-1 ship.
B5 (VerifierConfig altitude refactor) — best done once we have a second app under apps/ to prove the pluggability claim.

What I’m NOT doing without explicit ask

Running a Round-2 deep-research workflow (~3M tokens). Pass 1 produced enough to ship phase 1.
Pre-emptive vector-store benchmark (D2 requires us to host hardware or pay for a benchmark service).
Closed-tool reverse engineering (A1) — high research cost, low immediate yield.

Open an issue tagged research or architecture if you want any of these run.