Skip to content

Rebuild orchestration context inheritance on a stable provider contract #259

Closed as not planned

Description

@Juliusolsson05

Problem

The original orchestration context inheritance path is being removed from the active MCP behavior because it is too unstable for real child-agent work.

The old implementation attempted to duplicate or translate the parent provider transcript, resume the child from that cloned transcript, then rely on a bootstrap handoff marker and transcript cut points to separate inherited context from the child's actual output. In practice this was fragile:

  • child agents could retain stale parent identity or continue from the wrong conversational frame
  • provider resume behavior differed between Claude and Codex
  • stale parent answers could be surfaced as child outputs
  • reads needed brittle handoff-marker cut points to recover the child work product
  • delivery/readiness races made prompt-delivery accounting hard to trust

Current state

The removal PR disables inheritance for orchestration-created agents. Children now start from clean provider conversations, and parents must include required context explicitly in the prompt.

The old inheritParentContext schema field may remain tolerated for compatibility with stale MCP tool caches, but it should be ignored until this issue is resolved.

Goal

Bring back orchestration context inheritance only after there is a stable contract that proves a child can receive parent context without inheriting parent identity or corrupting output reads.

Acceptance criteria

  • Define a provider-neutral context handoff model that does not depend on resuming a child as if it were the parent.
  • Preserve separate child transcripts and never append child work to the parent transcript.
  • Make child output reads unambiguous without relying on fragile inherited transcript cut points.
  • Cover same-provider and cross-provider launches for Claude and Codex.
  • Include regression tests for nested orchestration, stale parent answers, failed inheritance, follow-up prompts, and closed/read output behavior.
  • Only re-enable inheritParentContext after the behavior is stable enough to expose through MCP again.

Activity

  1. Juliusolsson05 commented on Sep 27, 2026

    @Juliusolsson05
    OwnerAuthor

    Closing (owner decision, 2026-09-27). Child agents inheriting the parent's conversation stays off (disabled in #260). Passing context through the prompt or a brief file works well and is the supported approach. Reopen when: there's a concrete need that brief files can't meet, with a stable design.

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

    class:C7-new-featureBug cluster of a recently shipped featuretype:featureNew capability (out of quality-loop fix scope)

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions