96836e15ed
## Summary This PR updates the goal continuation prompt to address feedback from early adopters. There are two primary changes: 1. Goal continuation and budget-limit steering prompts now use hidden user-context messages instead of hidden developer messages. 2. The goal continuation prompt is refined to improve the model's ability to fully complete the active goal rather than stop at a smaller or merely passing subset. The user-message transition is important for two reasons. First, it eliminates an issue where older steering messages could be responded to again after a new turn. Second, it works better with compaction because user messages are treated differently from developer messages during compaction. The prompt refinements make persistence explicit, ground work in current evidence, encourage `update_plan` for multi-step progress visibility, and require stronger completion audits before calling `update_goal`. It also removes the elapsed-time reporting in the prompt; I saw evidence that this was causing the model to shortcut work as it became nervous about time. These changes were tested with evals. Chriss4123 has also been running independent evals in [#19910](https://github.com/openai/codex/issues/19910), and many of the improvements in this PR were suggested by him. ## Verification - Tested with evals. - Added and updated focused `codex-core` coverage for hidden goal user context, continuation and budget-limit request shape, prompt rendering, and objective delimiter escaping.
codex-core
This crate implements the business logic for Codex. It is designed to be used by the various Codex UIs written in Rust.
Dependencies
Note that codex-core makes some assumptions about certain helper utilities being available in the environment. Currently, this support matrix is:
macOS
Expects /usr/bin/sandbox-exec to be present.
When using the workspace-write sandbox policy, the Seatbelt profile allows
writes under the configured writable roots while keeping .git (directory or
pointer file), the resolved gitdir: target, and .codex read-only.
Network access and filesystem read/write roots are controlled by
SandboxPolicy. Seatbelt consumes the resolved policy and enforces it.
Seatbelt also keeps the legacy default preferences read access
(user-preference-read) needed for cfprefs-backed macOS behavior.
Linux
Expects the binary containing codex-core to run the equivalent of codex sandbox linux (legacy alias: codex debug landlock) when arg0 is codex-linux-sandbox. See the codex-arg0 crate for details.
Legacy SandboxPolicy / sandbox_mode configs are still supported on Linux.
They can continue to use the legacy Landlock path when the split filesystem
policy is sandbox-equivalent to the legacy model after cwd resolution.
Split filesystem policies that need direct FileSystemSandboxPolicy
enforcement, such as read-only or denied carveouts under a broader writable
root, automatically route through bubblewrap. The legacy Landlock path is used
only when the split filesystem policy round-trips through the legacy
SandboxPolicy model without changing semantics. That includes overlapping
cases like /repo = write, /repo/a = none, /repo/a/b = write, where the
more specific writable child must reopen under a denied parent.
The Linux sandbox helper prefers the first bwrap found on PATH outside the
current working directory whenever it is available. If bwrap is present but
too old to support --argv0, the helper keeps using system bubblewrap and
switches to a no---argv0 compatibility path for the inner re-exec. If
bwrap is missing, it falls back to the bundled codex-resources/bwrap
binary shipped with Codex and Codex surfaces a startup warning through its
normal notification path instead of printing directly from the sandbox helper.
Codex also surfaces a startup warning when bubblewrap cannot create user
namespaces. WSL2 uses the normal Linux bubblewrap path. WSL1 is not supported
for bubblewrap sandboxing because it cannot create the required user
namespaces, so Codex rejects sandboxed shell commands that would enter the
bubblewrap path before invoking bwrap.
Windows
Legacy SandboxPolicy / sandbox_mode configs are still supported on
Windows. Legacy read-only and workspace-write policies imply full
filesystem read access; exact readable roots are represented by split
filesystem policies instead.
The elevated Windows sandbox also supports:
- legacy
ReadOnlyandWorkspaceWritebehavior - split filesystem policies that need exact readable roots, exact writable roots, or extra read-only carveouts under writable roots
- backend-managed system read roots required for basic execution, such as
C:\Windows,C:\Program Files,C:\Program Files (x86), andC:\ProgramData, when a split filesystem policy requests platform defaults
The unelevated restricted-token backend still supports the legacy full-read
Windows model for legacy ReadOnly and WorkspaceWrite behavior. It also
supports a narrow split-filesystem subset: full-read split policies whose
writable roots still match the legacy WorkspaceWrite root set, but add extra
read-only carveouts under those writable roots.
New [permissions] / split filesystem policies remain supported on Windows
only when they can be enforced directly by the selected Windows backend or
round-trip through the legacy SandboxPolicy model without changing semantics.
Policies that would require direct explicit unreadable carveouts (none) or
reopened writable descendants under read-only carveouts still fail closed
instead of running with weaker enforcement.
All Platforms
Expects the binary containing codex-core to simulate the virtual
apply_patch CLI when arg1 is --codex-run-as-apply-patch. See the
codex-arg0 crate for details.