## Summary
- remove the remaining model-visible guardian-specific `on-request`
prompt additions so enabling the feature does not change the main
approval-policy instructions
- neutralize user-facing guardian wording to talk about automatic
approval review / approval requests rather than a second reviewer or
only sandbox escalations
- tighten guardian retry-context handling so agent-authored
`justification` stays in the structured action JSON and is not also
injected as raw retry context
- simplify guardian review plumbing in core by deleting dead
prompt-append paths and trimming some request/transcript setup code
## Notable Changes
- delete the dead `permissions/approval_policy/guardian.md` append path
and stop threading `guardian_approval_enabled` through model-facing
developer-instruction builders
- rename the experimental feature copy to `Automatic approval review`
and update the `/experimental` snapshot text accordingly
- make approval-review status strings generic across shell, patch,
network, and MCP review types
- forward real sandbox/network retry reasons for shell and unified-exec
guardian review, but do not pass agent-authored justification as raw
retry context
- simplify `guardian.rs` by removing the one-field request wrapper,
deduping reasoning-effort selection, and cleaning up transcript entry
collection
## Testing
- `just fmt`
- full validation left to CI
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- add the guardian reviewer flow for `on-request` approvals in command,
patch, sandbox-retry, and managed-network approval paths
- keep guardian behind `features.guardian_approval` instead of exposing
a public `approval_policy = guardian` mode
- route ordinary `OnRequest` approvals to the guardian subagent when the
feature is enabled, without changing the public approval-mode surface
## Public model
- public approval modes stay unchanged
- guardian is enabled via `features.guardian_approval`
- when that feature is on, `approval_policy = on-request` keeps the same
approval boundaries but sends those approval requests to the guardian
reviewer instead of the user
- `/experimental` only persists the feature flag; it does not rewrite
`approval_policy`
- CLI and app-server no longer expose a separate `guardian` approval
mode in this PR
## Guardian reviewer
- the reviewer runs as a normal subagent and reuses the existing
subagent/thread machinery
- it is locked to a read-only sandbox and `approval_policy = never`
- it does not inherit user/project exec-policy rules
- it prefers `gpt-5.4` when the current provider exposes it, otherwise
falls back to the parent turn's active model
- it fail-closes on timeout, startup failure, malformed output, or any
other review error
- it currently auto-approves only when `risk_score < 80`
## Review context and policy
- guardian mirrors `OnRequest` approval semantics rather than
introducing a separate approval policy
- explicit `require_escalated` requests follow the same approval surface
as `OnRequest`; the difference is only who reviews them
- managed-network allowlist misses that enter the approval flow are also
reviewed by guardian
- the review prompt includes bounded recent transcript history plus
recent tool call/result evidence
- transcript entries and planned-action strings are truncated with
explicit `<guardian_truncated ... />` markers so large payloads stay
bounded
- apply-patch reviews include the full patch content (without
duplicating the structured `changes` payload)
- the guardian request layout is snapshot-tested using the same
model-visible Responses request formatter used elsewhere in core
## Guardian network behavior
- the guardian subagent inherits the parent session's managed-network
allowlist when one exists, so it can use the same approved network
surface while reviewing
- exact session-scoped network approvals are copied into the guardian
session with protocol/port scope preserved
- those copied approvals are now seeded before the guardian's first turn
is submitted, so inherited approvals are available during any immediate
review-time checks
## Out of scope / follow-ups
- the sandbox-permission validation split was pulled into a separate PR
and is not part of this diff
- a future follow-up can enable `serde_json` preserve-order in
`codex-core` and then simplify the guardian action rendering further
---------
Co-authored-by: Codex <noreply@openai.com>
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
## Summary
- record a realtime close developer message when a new realtime session
replaces an active one
- assert the replacement marker through the mocked responses request
path
---------
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: Charles Cunningham <ccunningham@openai.com>
This updates the on-request permissions instructions so likely
sandbox-related network failures during dependency installation are
treated as escalation candidates.
Repro:
- Run `codex -a on-request -s workspace-write` in a fresh temp dir.
- Prompt: `Build a new rust app with one dependency, anyhow, and try
installing the dependency`.
- Before this change, DNS/registry failures like `Could not resolve
host: index.crates.io` could be treated like ordinary transient failures
and not escalate.
Fix:
- Clarify that likely sandbox-related network errors such as DNS/host
resolution, registry/index access, and dependency download failures
should trigger escalation.
Validation:
- Rebuild the CLI and rerun the same repro. The same instructions should
now be more likely to trigger escalation instead of silently stopping.
Related Slack canvas:
- https://openai.enterprise.slack.com/docs/T0BQTNSUF/F0ACVNJAV09
This PR replaces the old `additional_permissions.fs_read/fs_write` shape
with a shared `PermissionProfile`
model and wires it through the command approval, sandboxing, protocol,
and TUI layers. The schema is adopted from the
`SkillManifestPermissions`, which is also refactored to use this unified
struct. This helps us easily expose permission profiles in app
server/core as a follow-up.
## Summary
Introduces the initial implementation of Feature::RequestPermissions.
RequestPermissions allows the model to request that a command be run
inside the sandbox, with additional permissions, like writing to a
specific folder. Eventually this will include other rules as well, and
the ability to persist these permissions, but this PR is already quite
large - let's get the core flow working and go from there!
<img width="1279" height="541" alt="Screenshot 2026-02-15 at 2 26 22 PM"
src="https://github.com/user-attachments/assets/0ee3ec0f-02ec-4509-91a2-809ac80be368"
/>
## Testing
- [x] Added tests
- [x] Tested locally
- [x] Feature
## Summary
This feature is now reasonably stable, let's remove it so we can
simplify our upcoming iterations here.
## Testing
- [x] Existing tests pass
This removes overly directed language about how the model should behave
when it's in `approval_policy=never` mode.
---------
Co-authored-by: Dylan Hurd <dylan.hurd@openai.com>
## Summary
This PR consolidates base_instructions onto SessionMeta /
SessionConfiguration, so we ensure `base_instructions` is set once per
session and should be (mostly) immutable, unless:
- overridden by config on resume / fork
- sub-agent tasks, like review or collab
In a future PR, we should convert all references to `base_instructions`
to consistently used the typed struct, so it's less likely that we put
other strings there. See #9423. However, this PR is already quite
complex, so I'm deferring that to a follow-up.
## Testing
- [x] Added a resume test to assert that instructions are preserved. In
particular, `resume_switches_models_preserves_base_instructions` fails
against main.
Existing test coverage thats assert base instructions are preserved
across multiple requests in a session:
- Manual compact keeps baseline instructions:
core/tests/suite/compact.rs:199
- Auto-compact keeps baseline instructions:
core/tests/suite/compact.rs:1142
- Prompt caching reuses the same instructions across two requests:
core/tests/suite/prompt_caching.rs:150 and
core/tests/suite/prompt_caching.rs:157
- Prompt caching with explicit expected string across two requests:
core/tests/suite/prompt_caching.rs:213 and
core/tests/suite/prompt_caching.rs:222
- Resume with model switch keeps original instructions:
core/tests/suite/resume.rs:136
- Compact/resume/fork uses request 0 instructions for later expected
payloads: core/tests/suite/compact_resume_fork.rs:215
# External (non-OpenAI) Pull Request Requirements
Before opening this Pull Request, please read the dedicated
"Contributing" markdown file or your PR may be closed:
https://github.com/openai/codex/blob/main/docs/contributing.md
If your PR conforms to our contribution guidelines, replace this text
with a detailed and high quality description of your changes.
Include a link to a bug report or enhancement request.
- Add a single builder for developer permissions messaging that accepts
SandboxPolicy and approval policy. This builder now drives the developer
“permissions” message that’s injected at session start and any time
sandbox/approval settings change.
- Trim EnvironmentContext to only include cwd, writable roots, and
shell; removed sandbox/approval/network duplication and adjusted XML
serialization and tests accordingly.
Follow-up: adding a config value to replace the developer permissions
message for custom sandboxes.
Jack Mousseau ·
Charley Cunningham ·
Celia Chen ·
Ahmed Ibrahim ·
daniel-oai ·
Dylan Hurd ·
Fouad Matin ·