feat(analyst): a TraceAnalysisEngine a consumer can build without Python, so createTraceAnalyst is reachable

[drewstone]

At agent-eval 0.171.0 createTraceAnalyst(definition, options) requires options.engine: TraceAnalysisEngine, and that field is not optional (dist/default-registry-B0s2zU-s.d.ts:79).

Grepping the whole installed dist for a constructor returns exactly one:

$ grep -h "): TraceAnalysisEngine" dist/*.d.ts
dist/index-DT73JraI.d.ts:178  declare function createDspyRlmTraceEngine(options: DspyRlmTraceEngineOptions): TraceAnalysisEngine;

DspyRlmTraceEngineOptions needs runner?: ExternalOptimizerRunnerCommand ("Python command used to load agent-eval-rpc[dspy]. Default: python", index-DT73JraI.d.ts:172) plus a caller-owned ExternalOptimizerModelCall and a recordExecution sink.

Measured consequence, run against the installed packages:

buildDefaultAnalystRegistry().list()
  -> [{ id: "efficiency-behavioral", version: "2.0.0", cost: { kind: "deterministic" } }]

analystsFromRegistry(thatRegistry)     // @tangle-network/agent-runtime/kernel 0.185.2
  -> throws: analystsFromRegistry: "failure-mode", "intent-divergence", "knowledge-gap",
     "knowledge-poisoning", "improvement" are not registered;
     the registry has "efficiency-behavioral"

So a consumer that already has a model seam — an AgentProfile on a bridge or a sandbox seat, driven through agent-runtime streamAgentTurn — cannot build a registry that satisfies analystsFromRegistry, and cannot reach any of the five DEFAULT_TRACE_ANALYST_KINDS. The documented composition supervise(profile, task, { analysts: analystsFromRegistry(registry) }) (agent-runtime docs/canonical-api.md line 183) has no reachable path unless the consumer stands up Python.

Ask

Export a TraceAnalysisEngine constructor whose model seam is a plain call — an LlmCallRequest -> LlmCallResult function, or a ChatClient — so a consumer can supply its own executor. Something equivalent to:

declare function createCallTraceEngine(opts: {
  call: (req: LlmCallRequest) => Promise<LlmCallResult>
  recordExecution?: (row: unknown) => void
}): TraceAnalysisEngine

Then buildDefaultAnalystRegistry({ engine }) registers the five default kinds, analystsFromRegistry stops throwing, and a consumer deletes its hand-rolled lens.

Why we care, with the number

discovery-lab carries tools/analysts.mjs, 496 lines, whose createTraceBehaviorAnalyst hand-builds the { kinds, run } lens shape because of exactly this. About 150 of those lines are a prompt renderer over toolSpansToTraceAnalysisStore output that the engine would own. It is the last blocker on deleting the file.

Second, smaller gap in the same area

The one deterministic analyst that IS reachable, behavioralAnalyst, emits only monotonic-input-growth | output-length-decay | single-tool-dependency | no-self-verification (default-registry-B0s2zU-s.d.ts:7). There is no repeated-call / stuck-loop signal on the lens path. repeatedActionDetector has that invariant but is a StreamingDetector over pre-hashed fingerprints, not a store reader; computeToolUseMetrics and toolWasteView take a TraceStore while the lens contract hands a TraceAnalysisStore, and their duplicateRate is set membership, not a consecutive run. A deterministic, $0 loop panel over a TraceAnalysisStore would let a consumer delete another ~110 lines.

[drewstone]

PR #719 ships the constructor: createChatTraceEngine({ chat }), a TraceAnalysisEngine that runs in Node over a caller-owned ChatClient — no Python, no subprocess, no loopback proxy. buildDefaultAnalystRegistry({ engine }) then registers all five DEFAULT_TRACE_ANALYST_KINDS and analystsFromRegistry stops throwing; a test drives failure-mode end to end through registry.run() against a real OTLP store.

One shape change from the ask: it takes chat: ChatClient rather than a bare call function — the second form this issue named. A bare function cannot state maximumAttempts, and the priced ledger reservation needs it, since a transport that retries three times can bill three times. A caller holding a bare call adapts in one line: createChatClient({ transport: 'custom', chat: call, defaultModel, maximumAttempts }).

No closing keyword on the PR, because the second gap in this issue is not addressed: the deterministic, $0 repeated-call / stuck-loop panel over a TraceAnalysisStore. That is a separate change to behavioralAnalyst with its own claim about a consecutive run versus duplicateRate's set membership. Leaving this issue open for it.