## Summary
Some permission profiles can encode filesystem reads that should remain
unavailable to the agent. Before this change, the model-visible context
and automatic approval review prompt summarized the effective
permissions as a legacy sandbox mode, which can omit permission-profile
filesystem entries from escalation decisions.
For example, a profile can grant workspace access while denying a
private subtree across every workspace root:
```toml
default_permissions = "restricted-workspace"
[permissions.restricted-workspace.workspace_roots]
"/Users/alice/project" = true
"/Users/alice/other-project" = true
[permissions.restricted-workspace.filesystem]
":minimal" = "read"
[permissions.restricted-workspace.filesystem.":workspace_roots"]
"." = "write"
"private" = "deny"
"private/**" = "deny"
```
The context window now describes the workspace roots and effective
filesystem side of the `PermissionProfile` directly, with deny entries
marked as non-escalatable:
```xml
<environment_context>
<cwd>/Users/alice/project</cwd>
<shell>zsh</shell>
<filesystem><workspace_roots><root>/Users/alice/project</root><root>/Users/alice/other-project</root></workspace_roots><permission_profile type="managed"><file_system type="restricted"><entry access="read"><special>:minimal</special></entry><entry access="write"><path>/Users/alice/project</path></entry><entry access="write"><path>/Users/alice/other-project</path></entry><entry access="deny" escalatable="false"><path>/Users/alice/project/private</path></entry><entry access="deny" escalatable="false"><path>/Users/alice/other-project/private</path></entry><entry access="deny" escalatable="false"><glob>/Users/alice/project/private/**</glob></entry><entry access="deny" escalatable="false"><glob>/Users/alice/other-project/private/**</glob></entry></file_system></permission_profile></filesystem>
</environment_context>
```
Managed requirements can impose the same kind of deny-read restriction:
```toml
[permissions.filesystem]
deny_read = [
"/Users/alice/project/private",
"/Users/alice/project/private/**",
]
```
The automatic approval review prompt also receives the parent turn's
denied-read context, so review decisions can account for the active
permission profile.
## What Changed
- Render the effective filesystem profile in `<environment_context>`,
including profile type, filesystem entries, workspace roots, and
non-escalatable deny entries.
- Persist effective `workspace_roots` in `TurnContextItem` so
resumed/replayed context does not have to bind `:workspace_roots`
through legacy `cwd` fallback.
- Add explicit permission instructions that denied reads are policy
restrictions, not escalation targets.
- Pass the parent turn's denied-read context into automatic approval
reviews.
- Add targeted coverage for prompt rendering, workspace-root
materialization, replay context, and review prompt context.
- Keep the prompt-context test expectations platform-aware so the same
filesystem rendering assertions pass on Unix and Windows paths.
## Testing
- `just test -p codex-core
context::environment_context::tests::serialize_environment_context_with_full_filesystem_profile`
- `just test -p codex-core
context::environment_context::tests::turn_context_item_filesystem_uses_workspace_roots_instead_of_cwd`
- `just test -p codex-core
context::permissions_instructions::permissions_instructions_tests::builds_permissions_from_profile_with_denied_reads`
- `just fix -p codex-core`
I also attempted `just test -p codex-core`; the changed prompt-context
tests passed, but the full local run did not complete cleanly in this
sandboxed macOS environment due unrelated user-shell `CODEX_SANDBOX*`
expectations and integration-test timeouts.
## Why
Guardian reviews already emit analytics events, but we do not expose
aggregate OpenTelemetry metrics for review volume, latency, token usage,
or terminal outcomes. That makes it harder to monitor Guardian behavior
during rollouts and to compare review outcomes by source, action type,
session kind, model, and failure mode.
## What Changed
- Added Guardian review metric names for count, total duration, time to
first token, and token usage in `codex-rs/otel`.
- Added `core/src/guardian/metrics.rs` to convert
`GuardianReviewAnalyticsResult` into sanitized metric tags covering
decision, terminal status, failure reason, approval request source,
reviewed action, session kind, risk/outcome, model, reasoning effort,
and context/truncation state.
- Emitted the new metrics from `track_guardian_review` for each terminal
Guardian review result.
## Testing
- Added
`guardian_review_metrics_record_counts_durations_and_token_usage`, which
verifies the emitted count, duration, TTFT, token usage histograms, and
tag set through the in-memory metrics exporter.
Split from the Guardian prompt cache key change. This PR only updates
codex-rs/core/src/guardian/mod.rs. Validation was not run per request;
this branch is expected to rely on the companion split PRs.
## Summary
- remove the unreachable ARC monitor path from MCP tool approval
handling
- delete the unused ARC monitor module/tests and trim the orphaned
safety-monitor decision plumbing
- keep `always allow` approvals on the existing auto-approval
short-circuit without a dead monitor hop
## Testing
- `cargo test -p codex-core mcp_tool_call`
- `just fmt`
- `just fix -p codex-core`
- `git diff --check`
## Additional validation
- Attempted `cargo test -p codex-core`; the library test target passed,
then the integration target failed in this local environment.
- The narrower MCP-focused rerun passed its unit coverage and only hit
missing local `test_stdio_server` binaries in filtered integration
cases.
## Why
Guardian review selection was hard-coded in `core`, which worked for the
default OpenAI path but did not give provider implementations a way to
choose backend-specific reviewer model IDs. That matters for Amazon
Bedrock: guardian review should run through the Bedrock/Mantle provider
using Bedrock's `openai.gpt-5.4` model ID, instead of accidentally
selecting a reviewer model that implies the OpenAI backend.
## What Changed
- Added provider-owned approval review model selection via
`ModelProvider::approval_review_model_selection`.
- Moved the existing default selection policy into the provider
abstraction: prefer the requested reviewer model when it is available,
otherwise fall back to the active turn model, preferring `Low` reasoning
when supported.
- Added an Amazon Bedrock override that pins guardian review to
`openai.gpt-5.4` with `Low` reasoning.
## Why
Long-running turns can accumulate enough denied auto-review decisions to
trip the global short-circuit even when those denials are spread far
apart. The breaker should still stop genuinely bad loops, but it should
judge recent behavior instead of lifetime turn history.
## What changed
- Replaced the lifetime `10 total denials` threshold with `10 denials in
the last 50 reviews`.
- Kept the existing `3 consecutive denials` interrupt behavior
unchanged.
- Tracked recent auto-review outcomes in the circuit breaker and updated
the warning copy to report the rolling-window count.
- Renamed the new rolling-window coverage to `auto_review_*` test names.
- Added coverage that confirms older denials fall out of the 50-review
window and no longer trigger the breaker.
## Validation
- `just fmt`
- `cargo test -p codex-core guardian_rejection_circuit_breaker --lib`
- `cargo test -p codex-core auto_review_rejection_circuit_breaker --lib`
## Why
Auto-review can deny an action that the user later decides they want to
retry. Today there is no TUI surface for selecting a recent denial and
sending explicit approval context back into the session, so users have
to restate intent manually and the retry can be reviewed without the
original denied action context.
This adds a narrow TUI-driven path for approving a recent denied action
while still keeping the retry inside the normal auto-review flow.
## What Changed
- Added `/auto-review-denials` to open a picker of recent denied
auto-review actions.
- Added a small in-memory TUI store for the 10 most recent denied
auto-review events.
- Selecting a denial sends the structured denied event back through the
existing core/app-server op path.
- Core now injects a developer message containing the approved action
JSON rather than the full assessment event.
- Auto-review transcript collection now preserves this specific approval
developer message so follow-up review sessions can see the user approval
context.
- Added TUI snapshot/unit coverage for the picker and approval dispatch
path.
- Added core coverage for retaining the approval developer message in
the auto-review transcript.
## Verification
- `cargo test -p codex-core
collect_guardian_transcript_entries_keeps_manual_approval_developer_message`
- `cargo test -p codex-tui auto_review_denials`
- `cargo test -p codex-tui
approving_recent_denial_emits_structured_core_op_once`
## Notes
This intentionally keeps retries going through auto-review. The approval
signal is context for the exact previously denied action, not a blanket
bypass for similar future actions.
## Summary
Give guardian network-access reviews the command context that triggered
a managed-network approval. The prompt JSON now includes the originating
tool call id, tool name, command argv, cwd, sandbox permissions,
additional permissions, justification, and tty state when a single
active tool call can be attributed.
The implementation keeps the trigger shape canonical by serializing
`GuardianNetworkAccessTrigger` directly and lets each runtime build that
trigger from its `ToolCtx`. Non-guardian approval prompts avoid cloning
the full trigger payload.
## UX changes
Guardian network-access reviews now include a `trigger` object that
explains what command caused the network approval. Instead of seeing
only the requested host, the guardian reviewer can also see the
originating tool call, argv, working directory, sandbox mode,
justification, and tty state.
Example payload the guardian reviewer can see:
```json
{
"tool": "network_access",
"target": "https://api.github.com:443",
"host": "api.github.com",
"protocol": "https",
"port": 443,
"trigger": {
"callId": "call_abc123",
"toolName": "shell",
"command": ["gh", "api", "/repos/openai/codex/pulls/18197"],
"cwd": "/workspace/codex",
"sandboxPermissions": "require_escalated",
"justification": "Fetch PR metadata from GitHub.",
"tty": false
}
}
```
The network review itself remains scoped to the network decision:
`target_item_id` stays `null`. `trigger.callId` is attribution context
only, so clients can still distinguish network reviews from
item-targeted command reviews.
## Verification
- Added coverage for serializing network trigger context in guardian
approval JSON.
- Added regression coverage that network guardian reviews do not reuse
`trigger.callId` as `target_item_id`.
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
Short circuit the convo if auto-review hits too many denials
## Testing
- [x] Added unit tests
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
Allow guardian to skip other fields and output only
`{"outcome":"allow"}` when the command is low risk.
This change lets guardian reviews use a non-strict text format while
keeping the JSON schema itself as plain user-visible schema data, so
transport strictness is carried out-of-band instead of through a schema
marker key.
## What changed
- Add an explicit `output_schema_strict` flag to model prompts and pass
it into `codex-api` text formatting.
- Set guardian reviewer prompts to non-strict schema validation while
preserving strict-by-default behavior for normal callers.
- Update the guardian output contract so definitely-low-risk decisions
may return only `{"outcome":"allow"}`.
- Treat bare allow responses as low-risk approvals in the guardian
parser.
- Add tests and snapshots covering the non-strict guardian request and
optional guardian output fields.
## Verification
- `cargo test -p codex-core guardian::tests::guardian`
- `cargo test -p codex-core guardian::tests::`
- `cargo test -p codex-core client_common::tests::`
- `cargo test -p codex-protocol
user_input_serialization_includes_final_output_json_schema`
- `cargo test -p codex-api`
- `git diff --check`
Note: `cargo test -p codex-core` was also attempted, but this desktop
environment injects ambient config/proxy state that causes unrelated
config/session tests expecting pristine defaults to fail.
---------
Co-authored-by: Dylan Hurd <dylan.hurd@openai.com>
Co-authored-by: Codex <noreply@openai.com>
## Summary
When auto-review is enabled, it should handle request_permissions tool.
We'll need to clean up the UX but I'm planning to do that in a separate
pass
## Testing
- [x] Ran locally
<img width="893" height="396" alt="Screenshot 2026-04-17 at 1 16 13 PM"
src="https://github.com/user-attachments/assets/4c045c5f-1138-4c6c-ac6e-2cb6be4514d8"
/>
---------
Co-authored-by: Codex <noreply@openai.com>
## Why
Guardian review analytics needs a Rust event shape that matches the
backend schema while avoiding unnecessary PII exposure from reviewed
tool calls. This PR narrows the analytics payload to the fields we
intend to emit and keeps shared Guardian assessment enums in protocol
instead of duplicating equivalent analytics-only enums.
## What changed
- Uses protocol Guardian enums directly for `risk_level`,
`user_authorization`, `outcome`, and command source values.
- Removes high-risk reviewed-action fields from the analytics payload,
including raw commands, display strings, working directories, file
paths, network targets/hosts, justification text, retry reason, and
rationale text.
- Makes `target_item_id` and `tool_call_count` nullable so the Codex
event can represent cases where the app-server protocol or producer does
not have those values.
- Keeps lower-risk structured reviewed-action metadata such as sandbox
permissions, permission profile, `tty`, `execve` source/program, network
protocol/port, and MCP connector/tool labels.
- Adds an analytics reducer/client test covering `codex_guardian_review`
serialization with an optional `target_item_id` and absent removed
fields.
## Verification
- `cargo test -p codex-analytics
guardian_review_event_ingests_custom_fact_with_optional_target_item`
- `cargo fmt --check`
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/17692).
* #17696
* #17695
* #17693
* __->__ #17692
## Summary
This is the minimal client-side follow-up for the Codex Auto Review
model slug rollout. It updates the guardian reviewer preferred model
from `gpt-5.4` to `codex-auto-review`, so the client can rely on the
backend catalog + Statsig mapping instead of hardcoding the GPT-5.4
slug.
Context:
https://openai.slack.com/archives/C0AF9328RL0/p1775777479388369?thread_ts=1775773094.071629&cid=C0AF9328RL0
## Testing
- `cargo fmt --package codex-core --check`
- `cargo test -p codex-core guardian::`
- `bazel test --experimental_remote_downloader= --test_output=errors
//codex-rs/core:core-unit-tests --test_arg=guardian`
**Summary**
This PR treats Guardian timeouts as distinct from explicit denials in
the core approval paths.
Timeouts now return timeout-specific guidance instead of Guardian
policy-rejection messaging.
It updates the command, shell, network, and MCP approval flows and adds
focused test coverage.
## Description
This PR introduces `review_id` as the stable identifier for guardian
reviews and exposes it in app-server `item/autoApprovalReview/started`
and `item/autoApprovalReview/completed` events.
Internally, guardian rejection state is now keyed by `review_id` instead
of the reviewed tool item ID. `target_item_id` is still included when a
review maps to a concrete thread item, but it is no longer overloaded as
the review lifecycle identifier.
## Motivation
We'd like to give users the ability to preempt a guardian review while
it's running (approve or decline).
However, we can't implement the API that allows the user to override a
running guardian review because we didn't have a unique `review_id` per
guardian review. Using `target_item_id` is not correct since:
- with execve reviews, there can be multiple execve calls (and therefore
guardian reviews) per shell command
- with network policy reviews, there is no target item ID
The PR that actually implements user overrides will use `review_id` as
the stable identifier.
## Description
We reuse a guardian thread for a given user thread when we can. However,
we had always sent the full transcript history every time we made a
followup review request to an existing guardian thread.
This is especially bad for long guardian threads since we keep
re-appending old transcript entries instead of just what has changed.
The fix is to just send what's new.
**Caveat**: Whenever a thread is compacted or rolled back, we fall back
to sending the full transcript to guardian again since the thread's
history has been modified. However in the happy path we get a nice
optimization.
## Before
Initial guardian review sends the full parent transcript:
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please check the repo visibility and push the docs fix if needed.
[2] tool gh_repo_view call: {"repo":"openai/codex"}
[3] tool gh_repo_view result: repo visibility: public
[4] assistant: The repo is public; I now need approval to push the docs fix.
>>> TRANSCRIPT END
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
And a followup to the same guardian thread would send the full
transcript again (including items 1-4 we already sent):
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please check the repo visibility and push the docs fix if needed.
[2] tool gh_repo_view call: {"repo":"openai/codex"}
[3] tool gh_repo_view result: repo visibility: public
[4] assistant: The repo is public; I now need approval to push the docs fix.
[5] user: Please push the second docs fix too.
[6] assistant: I need approval for the second docs fix.
>>> TRANSCRIPT END
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
## After
Initial guardian review sends the full parent transcript (this is
unchanged):
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please check the repo visibility and push the docs fix if needed.
[2] tool gh_repo_view call: {"repo":"openai/codex"}
[3] tool gh_repo_view result: repo visibility: public
[4] assistant: The repo is public; I now need approval to push the docs fix.
>>> TRANSCRIPT END
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
But a followup now sends:
```
The following is the Codex agent history added since your last approval assessment. Continue the same review conversation...
>>> TRANSCRIPT DELTA START
[5] user: Please push the second docs fix too.
[6] assistant: I need approval for the second docs fix.
>>> TRANSCRIPT DELTA END
The Codex agent has requested the following next action:
>>> APPROVAL REQUEST START
...
>>> APPROVAL REQUEST END
```
## Description
This PR changes guardian transcript compaction so oversized
conversations no longer collapse into a nearly empty placeholder.
Before this change, if the retained user history alone exceeded the
message budget, guardian would replace the entire transcript with
`<transcript omitted to preserve budget for planned action>`!
That meant approvals, especially network approvals, could lose the
recent tool call and tool result that explained what guardian was
actually reviewing. Now we keep a compact but usable transcript instead
of dropping it all.
### Before
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
<transcript omitted to preserve budget for planned action>
>>> TRANSCRIPT END
Conversation transcript omitted due to size.
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
Retry reason:
Sandbox blocked outbound network access.
Assess the exact planned action below. Use read-only tool checks when local state matters.
Planned action JSON:
{
"tool": "network_access",
"target": "https://example.com:443",
"host": "example.com",
"protocol": "https",
"port": 443
}
>>> APPROVAL REQUEST END
```
### After
```
The following is the Codex agent history whose request action you are assessing...
>>> TRANSCRIPT START
[1] user: Please investigate why uploads to example.com are failing and retry if needed.
[8] user: If the request looks correct, go ahead and try again with network access.
[9] tool shell call: {"command":["curl","-X","POST","https://example.com/upload"],"cwd":"/repo"}
[10] tool shell result: sandbox blocked outbound network access
>>> TRANSCRIPT END
Some conversation entries were omitted.
The Codex agent has requested the following action:
>>> APPROVAL REQUEST START
Retry reason:
Sandbox blocked outbound network access.
Assess the exact planned action below. Use read-only tool checks when local state matters.
Planned action JSON:
{
"tool": "network_access",
"target": "https://example.com:443",
"host": "example.com",
"protocol": "https",
"port": 443
}
>>> APPROVAL REQUEST END
```
## Description
Previously the `action` field on `EventMsg::GuardianAssessment`, which
describes what Guardian is reviewing, was typed as an arbitrary JSON
blob. This PR cleans it up and defines a sum type representing all the
various actions that Guardian can review.
This is a breaking change (on purpose), which is fine because:
- the Codex app / VSCE does not actually use `action` at the moment
- the TUI code that consumes `action` is updated in this PR as well
- rollout files that serialized old `EventMsg::GuardianAssessment` will
just silently drop these guardian events
- the contract is defined as unstable, so other clients have a fair
warning :)
This will make things much easier for followup Guardian work.
## Why
The old guardian review payloads worked, but they pushed too much shape
knowledge into downstream consumers. The TUI had custom JSON parsing
logic for commands, patches, network requests, and MCP calls, and the
app-server protocol was effectively just passing through an opaque blob.
Typing this at the protocol boundary makes the contract clearer.
## Summary
- reuse a guardian subagent session across approvals so reviews keep a
stable prompt cache key and avoid one-shot startup overhead
- clear the guardian child history before each review so prior guardian
decisions do not leak into later approvals
- include the `smart_approvals` -> `guardian_approval` feature flag
rename in the same PR to minimize release latency on a very tight
timeline
- add regression coverage for prompt-cache-key reuse without
prior-review prompt bleed
## Request
- Bug/enhancement request: internal guardian prompt-cache and latency
improvement request
---------
Co-authored-by: Codex <noreply@openai.com>
Michael Bolin