Codex
Using Codex subagents for code review? Ask for a fresh context window, or they may just fork your chat
Matt Pocock says Codex often starts subagents as a fork of the parent conversation, so they inherit everything the main agent has already seen. For review and spec-checking work, that defeats the point. He updated his skills to ask for a fresh context window.
What changed
On Oct 29, Matt Pocock (@mattpocockuk) posted that Codex was biasing most of its subagents to inherit the parent agent's context window, effectively forking the conversation. He merged a change to his public skills repo (mattpocock/skills PR #1341) that adds one phrase to every place a skill spawns a subagent: start it "with a fresh context window". The PR touches eight skills, including code-review, implement-spec, research, and grilling, plus a repo rule so new skills follow it.
Who is affected
Anyone who uses Codex subagents for work that should be independent of the main thread: code review, checking an implementation against a spec, comparing multiple designs, or background research. A forked reviewer has already absorbed the main agent's reasoning and assumptions, so it is more likely to agree with them. People who use subagents only to split up parallel exploration are affected less.
What to do now
- When you ask for a reviewer or checker, say so explicitly, for example: "Spawn a subagent with a fresh context window. Give it only the diff and the spec, not this conversation."
- Put the same wording in AGENTS.md or in any skill that dispatches subagents, since Codex follows those instructions when it delegates.
- Because a fresh subagent knows only its brief, write the brief so it stands on its own: the goal, the files to look at, and what to return.
- Keep forking where shared context is the point. Tibo's Day 12 note said reusing subagent context helps usage last longer, so a fork can be cheaper. Use a fresh context when independence matters more than cost.
What is not confirmed
This is one developer's report, not an OpenAI statement. OpenAI's Codex subagents documentation does not describe when Codex forks versus starts a subagent fresh, and does not list a setting to control it, so how often forking happens may depend on the model and Codex version. The PR states that it deliberately leaves out harness-specific ways to turn forking off.
Sources: https://x.com/mattpocockuk/status/2115751897558614073 and https://github.com/mattpocock/skills/pull/1341
Sources
- ReportMatt Pocock (@mattpocockuk) on X: Codex subagents fork the parent context · Oct 29, 2026
- Officialmattpocock/skills PR #1341: start subagents with a fresh context window · Oct 29, 2026
Get updates like this every morning
- ① Email
- ② Card on Stripe
- ③ 7 days free
Then $2/month · cancel anytime in one click