Codex + Transcription
Codex runs the work loop and Transcription performs the selected operation. The Composer below preselects Codex, GPT-5.6 Sol, and this tool so you can test one real task with the same billing and recovery path as any AgentSky run.
How the combination works
Codex decides when to call Transcription.
Codex owns the coding-agent loop: it explores the repository, changes files, runs commands, and returns evidence from the cloud workspace.
The current adapter uses Fish Audio transcribe-1. Supply audio as an accessible URL or a local file through the CLI, and use automatic language detection or a language code. The returned transcript can feed a separate summarization task, such as extracting decisions and action items from an interview.
Codex
Owns the task loop, context, model calls, tool choice, cloud computer, progress and final result.
Transcription
Performs the operation exposed by the current Transcription capability. Its input limits and output shape come from the live tool guide and adapter.
AgentSky
Keeps the selected agent, model, capability grant, task state and usage record together. Tool charges and model or runtime charges remain itemized separately.
Setup and first task
Start with one bounded transcription task.
The Composer carries Codex, GPT-5.6 Sol, and Transcription. Begin with the smallest input that lets you inspect the output before expanding the workflow.
Open the Codex Composer
The Composer above preselects Codex, GPT-5.6 Sol, and Transcription. Add the target, required inputs and a concrete expected result; the draft stays in place through sign-in.
State the Transcription boundary
Transcribe this recording faithfully. Flag unclear passages and preserve names and numbers without guessing.
Run and verify the returned result
Check the output against the source or brief, confirm that the result has the expected format, and inspect the task usage for model, runtime and tool charges.
Inputs, cost and limits
What Codex + Transcription establishes
Software engineering tasks that benefit from Codex's repository navigation, command execution, and iterative implementation loop.
Configure and verify
- Codex and GPT-5.6 Sol are preselected for the browser evaluation.
- Transcription is enabled from the current capability catalog with its live input and output guidance.
- The first task can be checked from the returned artifact, text or provider result before a larger workflow is attempted.
Still depends on the task
- A tool guide does not guarantee a provider result for malformed, blocked or unsupported Transcription inputs.
- A successful tool call does not establish that the result is correct for your business decision; inspect the output at its source boundary.
- Total task cost can include model tokens, active computer time and the selected tool's billable units.
Current Transcription rates
fish-audio.transcribe
$0.0002 per second.
Tool-specific details
Use the current Transcription contract.
The current adapter uses Fish Audio transcribe-1. Supply audio as an accessible URL or a local file through the CLI, and use automatic language detection or a language code. The returned transcript can feed a separate summarization task, such as extracting decisions and action items from an interview.
Audio input
The adapter accepts an audio URL or base64 bytes. The CLI reads local files into base64; the configured audio size limit is 50,000,000 bytes.
Language and result
language defaults to auto. The adapter returns text, detected or requested language, and provider.
Timestamps and subtitles
The adapter explicitly sends ignore_timestamps: true. It does not produce SRT or speaker labels, even though the CLI catalog lists a format option.
Common questions