Wayfinder: Define Expert Talk for Web and TUI #251

Destination

A reviewed, decision-complete specification for Expert Talk in Pythinker Web and the interactive TUI. A user selects two configured models, runs one bounded structured exchange, and receives one fused answer.

Status: complete. All nine decision children are resolved and closed. Implementation tickets can now be written without making new V1 product, protocol, safety, recovery, budget, client-contract, experience, or rollout decisions.

Expert Talk V1

Expert Talk combines two configured models instead of choosing between them. Both independently analyze the same request, review the other's complete reasoning, revise their positions, and a fresh inference using the configured Fusion Lead binding critically merges the exchange into one authoritative answer.

The successful protocol is:

2 independent openings
→ 2 attempted reciprocal reviews, at least 1 required
→ 1 fresh Fusion Lead inference
→ 1 direct assistant answer

Participants are run-scoped, read-only, context-bounded, and never restored as live sessions.

Resolved decisions

• Activation and session semantics: one-shot next-turn arm, durable session-local pair, immutable run snapshot, exclusive turn control, no restored or forked live execution.
• Model eligibility and routing guarantees: two ordered canonical configured IDs, authoritative pre-acceptance routing, eligible and distinct effective bindings, coherent routing revision, immutable execution provenance, no silent fallback.
• Conversation and fusion protocol: parallel independent openings, one reciprocal review round, fresh Lead-model critical fusion, direct answer plus separate fusion notes, no third judge or open loop.
• Permissions and trust boundaries: only Read, Glob, Grep, and ReadMediaFile; no shell, writes, network tools, delegation, or hidden authority; peer and repository text remain untrusted data.
• Context and request budgets: fixed output/tool/request caps, conservative worst-case admission, complete-packet runtime rechecks, 5–12 successful model requests, at most 24 provider attempts, no silent truncation or spill path.
• Failure, cancellation, and recovery: both openings required, one review sufficient, no fusion fallback, ten-minute run ceiling, fan-out cancellation, immutable partial artifacts, crash terminalization, full new-run retry.
• API, transcript, and compatibility contracts: session authority, versioned status/run/result schemas, snapshot-first dedicated events, stable errors, normal user-plus-fused-answer transcript, readable old-client results, no unknown events sent to old clients.
• Web and TUI experience: the Fusion Harness visual system maps to Expert Talk roles, with pair setup, one-shot widget, responsive agent grid, full-width fusion row, partial/failure states, accessible controls, and a separate expandable exchange. Open the prototype.
• Rollout and acceptance evidence: default-off expert_talk flag, shared backend status, explicit cost/privacy disclosure, focused and live-provider proof, generated Web checks, one minor published-package changeset, and a separate reviewed default-on decision.

Notes

• Domain: Expert Talk spans agent-core-v2, the shared protocol and gateway, klient, transcript, Pythinker Web, and the interactive TUI.
• This map resolves decisions only. It does not claim feature implementation, live-provider evidence, delivery, release, or default-on activation.
• The public prototype source is preserved in this gist.
• Historical run records are explained by their frozen bindings and schemas, never reinterpreted through current catalog or routing state.

Not yet specified

Nothing required for Expert Talk V1 implementation. Any new scope must become a separate enhancement rather than an implicit V1 behavior.

Out of scope

• Feature implementation and delivery work.
• Legacy engine-v1 and non-interactive print-mode support in the first delivery.
• More than two participants, a third judge, multiple review rounds, or persistent participant sessions.
• Participant write access or any bypass of existing model routing, workspace, secret, or tool policy.