Skip to content

fix(hints): parse current Codex user-message event shapes in rollout renderer #842

Description

@ScriptedAlchemy

Evidence

A 12-hour hook-hint audit found six canonical user.texthooks.additional_context pairs, but the renderer produced no submitted prompts:

  • 334 model messages observed
  • submitted_user_messages=0
  • developer_trace_hints=0
  • rendered_cases=0

The current Codex rollouts encode prompts as response_item user.text and event_msg.item_completed UserMessage. scripts/render-codex-hook-inputs.py recognizes only the older shape. The same legacy-only assumption appears in crates/tracedecay-sessions/src/runtime/hosts/codex.rs.

This makes hint evaluation falsely look like an empty corpus and prevents safe classifier tuning.

Expected behavior

Canonical current Codex user messages are recovered through the maintained session/redaction authority and paired with their hook context without fabricating or duplicating prompts.

Acceptance

  • Add real-rollout fixtures for both current event shapes.
  • Render the six known 12-hour prompt/context pairs from the bounded audit.
  • Preserve legacy parsing and canonical redaction/content authority.
  • Prove malformed, duplicate, and non-user events are rejected or ignored deterministically.
  • Re-run the deterministic hook-hint evals and report scored positives/silence; do not tune classifiers from an empty-positive sample.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't working

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions