conduit-sdk（conduit SDK 命令行）

Role

• Treat this skill as the canonical public Conduit integration guide for agents.
• Optimize for first successful integration, not exhaustive SDK tours.
• Use only the documented surfaces in the bundled references. If a method, type, or request shape is not shown there, say so and do not invent it.

How to load references

1. Read references/common.md first.
2. Detect the target language before giving code:

- Explicit user language or framework wins. - Otherwise inspect repo manifests near the task.

1. Read only one of references/typescript.md, references/python.md, references/go.md, or references/rust.md.
2. Read multiple language references only for ports, comparisons, or polyglot repos.
3. If the repo still points to multiple equally plausible SDK languages after inspection, ask one targeted language question.

Framework and workflow defaults

• If the repo or request implies a framework, answer in that framework: Next.js for TypeScript, FastAPI for Python, net/http for Go, axum for Rust.
• Otherwise stay framework-agnostic and use language-native SDK examples.
• Default to reports unless the user explicitly asks for target-versus-group comparison.
• First answers should show one obvious path: upload media, create a report from media_id, target the speaker with magic_hint, add the webhook handler, verify the signature on the raw body, dedupe on event.id, fetch the completed report.
• Prefer webhooks as the default completion path in production. Conduit jobs are asynchronous and often take around 150 seconds.
• Use receipt handles like wait() or stream() only for local scripts, operator tooling, or explicit synchronous control.
• Reach for primitives.media, primitives.jobs, and primitives.entities only when the user needs upload reuse, job inspection or cancellation, or stable entity management.
• When the user needs to pick one person from uploaded media, prefer magic_hint first with a concrete hint like the interviewed person.

Conduit-specific guardrails

• Verify webhook signatures on the exact raw request body before parsing the event.
• Deduplicate webhook deliveries on event.id before doing side effects.
• Keep onboarding examples on the stable surface unless the user explicitly asks for lower-level control.
• Preserve idempotency keys when the caller already has a stable external job identifier.
• Match package names, imports, and method names exactly from the chosen language reference.
• If the user asks for browser TypeScript, call out the dangerouslyAllowBrowser requirement.

First-time integration checklist

1. Initialize the client with the correct package import and API key shape.
2. Upload the recording and reuse the returned media_id.
3. Create a reports job with target = magic_hint("the interviewed person").
4. Add the webhook route in the detected framework.
5. Verify the signature on the exact raw request body before parsing.
6. Deduplicate on event.id.
7. Handle report.completed and fetch the final report.
8. Mention wait() or stream() only as local tooling helpers when relevant.

When to branch out

• Use matching only for explicit target-versus-group comparisons in a named context.
• Use primitives.media when the recording should be uploaded once and reused by media_id.
• Use primitives.jobs for job inspection or cancellation.
• Use primitives.entities for stable entity management or backoffice flows.
• Use direct entity_id targeting only when the caller already knows the stable speaker identity.

Output expectations

• Keep first examples short: upload media, create a report from media_id, target with magic_hint, verify a webhook, dedupe, fetch the final report.
• Use the user language and framework idioms from the chosen reference instead of translating through TypeScript first.
• If the user is migrating between SDKs, map equivalent surfaces directly instead of re-explaining Conduit from scratch.
• Do not lead with matching, primitives, polling, or architecture talk when the user just needs the first integration working.
• If you are tuning this skill itself, use assets/trigger-queries.json as the trigger eval fixture set.