Skip to content

feat: Make recorded sessions richer and shareable #1920

Open
0 of 6 issues completed
Open

Description

@michaljacovi

Motivation

Cortex already records every exchange a coding agent makes: raw request and
response bytes, every retry attempt, TTFB and duration, all egress. Two uses
need more than that today:

  • Realistic replay benchmarks. Replay tools such as Inference-Perf need
    real sessions with their content, their structure (which calls belong to
    one user prompt, which are subagents) and their timing.
  • Sharing real sessions. Some users, starting with our team, want to
    share recorded sessions, for example with an evaluation group. That needs
    redaction, a way to keep a session out, and a format the receiver can load.

This builds on the import (#1856, #1876) and the content-free export (#1810,
#1854, #1812), and does not replace them.

Work

Suggested order: S1 first, because S5 depends on it. S2, S3 and S4 are
independent.

Agreed so far (2026-11-17)

  • The stored archive stays raw. Redaction runs by default on export, with
    redact-at-save as an option (S1).
  • Opt-out: an env var at launch that becomes a request header, plus an
    optional mid-session /no-record (S3).
  • Storage stays the block-dedup archive. Replay formats are generated on
    export, compressed (S5).

Still open: whether telemetry may join a live session later. We start by
joining at export time (S4).

Assisted-By: Claude (Anthropic AI) noreply@anthropic.com

Metadata

Metadata

Assignees

Labels

enhancementNew feature or request

Type

No type

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions