AssemblyAI Failure Triage
Overview
Analyze and triage failures by contract layer without exposing customer speech. Treat live data, credentials, spend, and destructive state as separately governed boundaries.
Prerequisites
- The target repository or integration path and the requested operator outcome.
- The AssemblyAI project, environment, region, data classification, and accountable owner.
- Current first-party documentation plus credentials only for a narrowly approved live check.
Current Contract
Transport status, transcript terminal status, WebSocket messages or close codes, webhook response behavior, and gateway results are separate failure layers. A 429 can represent shared account request or concurrency pressure; more keys do not multiply account capacity.
Authentication
For live work, inject ASSEMBLYAI_API_KEY from an approved secret manager and send the raw value only in the AssemblyAI Authorization header to the configured first-party host. Never print, commit, place in a URL, or expose it to an untrusted client. Callback secrets and temporary streaming tokens are separate credentials.
Instructions
- Capture environment, region, operation, safe ID, timestamp, and last state.
- Separate network, auth, validation, quota, processing, streaming, callback, and gateway failures.
- Reproduce with synthetic input whenever possible.
- Compare model and parameters to current mode-specific docs.
- Collect only safe status, code, timing, and request metadata.
- Classify retry, input correction, capacity action, or escalation.
Tool Discipline
Use Read, Glob, and Grep to inspect repository code, configuration, fixtures, and evidence. Use Write and Edit only for approved implementation or documentation changes. Do not call AssemblyAI, upload audio, open a streaming session, mint a token, replay a callback, deploy, rotate a key, or delete a transcript merely because this skill was invoked.
Approval Boundaries
Require an accountable owner before live audio processing, production credential or endpoint changes, paid model or capacity changes, content retention, callback replay, deployment, or deletion. Read-only repository inspection and synthetic offline validation do not authorize live vendor actions.
Failure Modes
- Repeated 401 responses need credential-owner review, not logging the key.
- Terminal transcript errors are not fixed by infinite polling.
- Streaming reconnects must not overwrite the prior session.
Output
Return the operation scope, environment, region, contract surface, authorization class, model and feature decisions, deterministic validation results, content-free identifiers, risks, cleanup or rollback state, and a concise pass/fail receipt. Exclude credentials, signed URLs, audio, transcript text, prompts, and customer-derived content.
Example
- Start with the named environment, approved regional host, synthetic fixture identity, and bounded operation budget.
- Finish with safe IDs, contract and assertion counts, terminal state, cleanup status, and the decision owner; never reproduce speech content.
Validation
Rerun the smallest relevant deterministic check, compare actual state with the requested outcome and current first-party contract, verify sensitive fields are absent from evidence, and confirm rollback, termination, or deletion state before reporting success.
References
Review the dated first-party evidence map before relying on any model, parameter, limit, price, region, or lifecycle claim.