- Add a hidden `realtime_conversation` feature flag and `/realtime`
slash command for start/stop live voice sessions.
- Reuse transcription composer/footer UI for live metering, stream mic
audio, play assistant audio, render realtime user text events, and
force-close on feature disable.
---------
Co-authored-by: Codex <noreply@openai.com>
Summary
- detect skill-invoking shell commands based on the original command
string, request approvals when needed, and cache positive decisions per
session
- keep implicit skill invocation emitted after approval and keep skill
approval decline messaging centralized to the shell handler
- expand and adjust skill approval tests to cover shell-based skill
scripts while matching the new detection expectations
Testing
- Not run (not requested)
#### What
Try matching `\w+`-namespaced model after `longest prefix` as heuristic
to match `ModelInfo` from list of candidates.
This shouldn't regress existing behavior:
- `gpt-5.2-codex` -> `gpt-5.2` if `gpt-5.2-codex` not present
- `gpt-5.3` -> `gpt-5` if `gpt-5.3` not present
- `gpt-9` still doesn't match anything
while being more forgiving for custom prefixes:
- `oai/gpt-5.3-codex` -> `gpt-5.3-codex`
#### Tests
Added unit test.
## Why
This PR switches the `shell_command` zsh-fork path over to
`codex-shell-escalation` so the new shell tool can use the shared
exec-wrapper/escalation protocol instead of the `zsh_exec_bridge`
implementation that was introduced in
https://github.com/openai/codex/pull/12052. `zsh_exec_bridge` relied on
UNIX domain sockets, which is not as tamper-proof as the FD-based
approach in `codex-shell-escalation`.
## What Changed
- Added a Unix zsh-fork runtime adapter in `core`
(`core/src/tools/runtimes/shell/unix_escalation.rs`) that:
- runs zsh-fork commands through
`codex_shell_escalation::run_escalate_server`
- bridges exec-policy / approval decisions into `ShellActionProvider`
- executes escalated commands via a `ShellCommandExecutor` that calls
`process_exec_tool_call`
- Updated `ShellRuntime` / `ShellCommandHandler` / tool spec wiring to
select a `shell_command` backend (`classic` vs `zsh-fork`) while leaving
the generic `shell` tool path unchanged.
- Removed the `zsh_exec_bridge`-based session service and deleted
`core/src/zsh_exec_bridge/mod.rs`.
- Moved exec-wrapper entrypoint dispatch to `arg0` by handling the
`codex-execve-wrapper` arg0 alias there, and removed the old
`codex_core::maybe_run_zsh_exec_wrapper_mode()` hooks from `cli` and
`app-server` mains.
- Added the needed `codex-shell-escalation` dependencies for `core` and
`arg0`.
## Tests
- `cargo test -p codex-core
shell_zsh_fork_prefers_shell_command_over_unified_exec`
- `cargo test -p codex-app-server turn_start_shell_zsh_fork --
--nocapture`
- verifies zsh-fork command execution and approval flows through the new
backend
- includes subcommand approve/decline coverage using the shared zsh
DotSlash fixture in `app-server/tests/suite/zsh`
- To test manually, I added the following to `~/.codex/config.toml`:
```toml
zsh_path = "/Users/mbolin/code/codex3/codex-rs/app-server/tests/suite/zsh"
[features]
shell_zsh_fork = true
```
Then I ran `just c` to run the dev build of Codex with these changes and
sent it the message:
```
run `echo $0`
```
And it replied with:
```
echo $0 printed:
/Users/mbolin/code/codex3/codex-rs/app-server/tests/suite/zsh
In this tool context, $0 reflects the script path used to invoke the shell, not just zsh.
```
so the tool appears to be wired up correctly.
## Notes
- The zsh subcommand-decline integration test now uses `rm` under a
`WorkspaceWrite` sandbox. The previous `/usr/bin/true` scenario is
auto-allowed by the new `shell-escalation` policy path, which no longer
produces subcommand approval prompts.
## 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
Send a request with `generate: falls` but a full set of tools and
instructions to pre-warm inference.
---------
Co-authored-by: Codex <noreply@openai.com>
## Summary
- tighten the memory-use decision boundary so agents skip memory only
for clearly self-contained asks
- make the quick memory pass more explicit and bounded (including a
lightweight search budget)
- add structured `<memory_citation>` requirements and examples for final
replies
- clarify memory update guidance and end-state wording for memory lookup
## Why
The previous template was directionally correct, but still left room for
inconsistent memory lookup behavior and citation formatting. This change
makes the default behavior, quick-pass scope, and citation output
contract much more explicit.
## Testing
- not run (prompt/template text change only)
Co-authored-by: jif-oai <jif@openai.com>
rm `PRESETS` list harcoded in `model_presets` as we now have bundled
`models.json` with equivalent info.
update logic to rely on bundled models instead, update tests.
- use `skills_for_cwd` lookup to scope allowed skills and build
invocation context for downstream processing
- add detection in `stream_events_utils` to classify tool calls as
implicit skill invocations per the proposal (script runners, extensions,
`scripts` dirs, and SKILL.md reads)
- deduplicate invocations per turn and emit analytics/OTEL events on the
same background queue as explicit invokes
## Summary
Persist network approval allow/deny decisions as `network_rule(...)`
entries in execpolicy (not proxy config)
It adds `network_rule` parsing + append support in `codex-execpolicy`,
including `decision="prompt"` (parse-only; not compiled into proxy
allow/deny lists)
- compile execpolicy network rules into proxy allow/deny lists and
update the live proxy state on approval
- preserve requirements execpolicy `network_rule(...)` entries when
merging with file-based execpolicy
- reject broad wildcard hosts (for example `*`) for persisted
`network_rule(...)`
## Why
Tool handlers and runtimes needed to pass the same turn/session context
for shell and non-shell workflows without duplicative ownership churn.
Using shared pointers avoids temporary lifetimes and keeps existing
behavior unchanged while simplifying call sites.
## What changed
- Converted `ToolCtx` to store shared context handles (`Arc`-based),
including updates across shell, apply-patch, and unified-exec paths.
- Updated orchestrator/runtime call sites to consume the shared context
consistently and remove brittle move/borrow patterns.
- Kept behavior unchanged while preparing the type surface for the new
shell escalation integration in the next stack commit.
## Verification
- Validated this commit stack point with `just clippy` and confirmed
workspace compiles cleanly in this stack state.
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/12583).
* #12584
* __->__ #12583
* #12556
## Why
The zsh integration tests were still brittle in two ways:
- they relied on `CODEX_TEST_ZSH_PATH` / environment-specific setup, so
they often did not exercise the patched zsh fork that `shell-tool-mcp`
ships
- once the tests consistently used the vendored zsh fork, they exposed
real Linux-specific zsh-fork issues in CI
In particular, the Linux failures were not just test noise:
- the zsh-fork launch path was dropping `ExecRequest.arg0`, so Linux
`codex-linux-sandbox` arg0 dispatch did not run and zsh wrapper-mode
could receive malformed arguments
- the
`turn_start_shell_zsh_fork_subcommand_decline_marks_parent_declined_v2`
test uses the zsh exec bridge (which talks to the parent over a Unix
socket), but Linux restricted sandbox seccomp denies `connect(2)`,
causing timeouts on `ubuntu-24.04` x86/arm
This PR makes the zsh tests consistently run against the intended
vendored zsh fork and fixes/hardens the zsh-fork path so the Linux CI
signal is meaningful.
## What Changed
- Added a single shared test-only DotSlash file for the patched zsh fork
at `codex-rs/exec-server/tests/suite/zsh` (analogous to the existing
`bash` test resource).
- Updated both app-server and exec-server zsh tests to use that shared
DotSlash zsh (no duplicate zsh DotSlash file, no `CODEX_TEST_ZSH_PATH`
dependency).
- Updated the app-server zsh-fork test helper to resolve the shared
DotSlash zsh and avoid silently falling back to host zsh.
- Kept the app-server zsh-fork tests configured via `config.toml`, using
a test wrapper path where needed to force `zsh -df` (and rewrite `-lc`
to `-c`) for the subcommand-decline test.
- Hardened the app-server subcommand-decline zsh-fork test for CI
variability:
- tolerate an extra `/responses` POST with a no-op mock response
- tolerate non-target approval ordering while remaining strict on the
two `/usr/bin/true` approvals and decline behavior
- use `DangerFullAccess` on Linux for this one test because it validates
zsh approval flow, not Linux sandbox socket restrictions
- Fixed zsh-fork process launching on Linux by preserving `req.arg0` in
`ZshExecBridge::execute_shell_request(...)` so `codex-linux-sandbox`
arg0 dispatch continues to work.
- Moved `maybe_run_zsh_exec_wrapper_mode()` under
`arg0_dispatch_or_else(...)` in `app-server` and `cli` so wrapper-mode
handling coexists correctly with arg0-dispatched helper modes.
- Consolidated duplicated `dotslash -- fetch` resolution logic into
shared test support (`core/tests/common/lib.rs`).
- Updated `codex-rs/exec-server/tests/suite/accept_elicitation.rs` to
use DotSlash zsh and hardened the zsh elicitation test for Bazel/zsh
differences by:
- resolving an absolute `git` path
- running `git init --quiet .`
- asserting success / `.git` creation instead of relying on banner text
## Verification
- `cargo test -p codex-app-server turn_start_zsh_fork -- --nocapture`
- `cargo test -p codex-exec-server accept_elicitation -- --nocapture`
- `bazel test //codex-rs/exec-server:exec-server-all-test
--test_output=streamed --test_arg=--nocapture
--test_arg=accept_elicitation_for_prompt_rule_with_zsh`
- CI (`rust-ci`) on the final cleaned commit: `Tests — ubuntu-24.04 -
x86_64-unknown-linux-gnu` and `Tests — ubuntu-24.04-arm -
aarch64-unknown-linux-gnu` passed in [run
22291424358](https://github.com/openai/codex/actions/runs/22291424358)
## Summary
Adds syntax highlighting to the TUI for fenced code blocks in markdown
responses and file diffs, plus a `/theme` command with live preview and
persistent theme selection. Uses syntect (~250 grammars, 32 bundled
themes, ~1 MB binary cost) — the same engine behind `bat`, `delta`, and
`xi-editor`. Includes guardrails for large inputs, graceful fallback to
plain text, and SSH-aware clipboard integration for the `/copy` command.
<img width="1554" height="1014" alt="image"
src="https://github.com/user-attachments/assets/38737a79-8717-4715-b857-94cf1ba59b85"
/>
<img width="2354" height="1374" alt="image"
src="https://github.com/user-attachments/assets/25d30a00-c487-4af8-9cb6-63b0695a4be7"
/>
## Problem
Code blocks in the TUI (markdown responses and file diffs) render
without syntax highlighting, making it hard to scan code at a glance.
Users also have no way to pick a color theme that matches their terminal
aesthetic.
## Mental model
The highlighting system has three layers:
1. **Syntax engine** (`render::highlight`) -- a thin wrapper around
syntect + two-face. It owns a process-global `SyntaxSet` (~250 grammars)
and a `RwLock<Theme>` that can be swapped at runtime. All public entry
points accept `(code, lang)` and return ratatui `Span`/`Line` vectors or
`None` when the language is unrecognized or the input exceeds safety
guardrails.
2. **Rendering consumers** -- `markdown_render` feeds fenced code blocks
through the engine; `diff_render` highlights Add/Delete content as a
whole file and Update hunks per-hunk (preserving parser state across
hunk lines). Both callers fall back to plain unstyled text when the
engine returns `None`.
3. **Theme lifecycle** -- at startup the config's `tui.theme` is
resolved to a syntect `Theme` via `set_theme_override`. At runtime the
`/theme` picker calls `set_syntax_theme` to swap themes live; on cancel
it restores the snapshot taken at open. On confirm it persists `[tui]
theme = "..."` to config.toml.
## Non-goals
- Inline diff highlighting (word-level change detection within a line).
- Semantic / LSP-backed highlighting.
- Theme authoring tooling; users supply standard `.tmTheme` files.
## Tradeoffs
| Decision | Upside | Downside |
| ------------------------------------------------ |
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
|
-----------------------------------------------------------------------------------------------------------------------
|
| syntect over tree-sitter / arborium | ~1 MB binary increase for ~250
grammars + 32 themes; battle-tested crate powering widely-used tools
(`bat`, `delta`, `xi-editor`). tree-sitter would add ~12 MB for 20-30
languages or ~35 MB for full coverage. | Regex-based; less structurally
accurate than tree-sitter for some languages (e.g. language injections
like JS-in-HTML). |
| Global `RwLock<Theme>` | Enables live `/theme` preview without
threading Theme through every call site | Lock contention risk
(mitigated: reads vastly outnumber writes, single UI thread) |
| Skip background / italic / underline from themes | Terminal BG
preserved, avoids ugly rendering on some themes | Themes that rely on
these properties lose fidelity |
| Guardrails: 512 KB / 10k lines | Prevents pathological stalls on huge
diffs or pastes | Very large files render without color |
## Architecture
```
config.toml ─[tui.theme]─> set_theme_override() ─> THEME (RwLock)
│
┌───────────────────────────────────────────┘
│
markdown_render ─── highlight_code_to_lines(code, lang) ─> Vec<Line>
diff_render ─── highlight_code_to_styled_spans(code, lang) ─> Option<Vec<Vec<Span>>>
│
│ (None ⇒ plain text fallback)
│
/theme picker ─── set_syntax_theme(theme) // live preview swap
─── current_syntax_theme() // snapshot for cancel
─── resolve_theme_by_name(name) // lookup by kebab-case
```
Key files:
- `tui/src/render/highlight.rs` -- engine, theme management, guardrails
- `tui/src/diff_render.rs` -- syntax-aware diff line wrapping
- `tui/src/theme_picker.rs` -- `/theme` command builder
- `tui/src/bottom_pane/list_selection_view.rs` -- side content panel,
callbacks
- `core/src/config/types.rs` -- `Tui::theme` field
- `core/src/config/edit.rs` -- `syntax_theme_edit()` helper
## Observability
- `tracing::warn` when a configured theme name cannot be resolved.
- `Config::startup_warnings` surfaces the same message as a TUI banner.
- `tracing::error` when persisting theme selection fails.
## Tests
- Unit tests in `highlight.rs`: language coverage, fallback behavior,
CRLF stripping, style conversion, guardrail enforcement, theme name
mapping exhaustiveness.
- Unit tests in `diff_render.rs`: snapshot gallery at multiple terminal
sizes (80x24, 94x35, 120x40), syntax-highlighted wrapping, large-diff
guardrail, rename-to-different-extension highlighting, parser state
preservation across hunk lines.
- Unit tests in `theme_picker.rs`: preview rendering (wide + narrow),
dim overlay on deletions, subtitle truncation, cancel-restore, fallback
for unavailable configured theme.
- Unit tests in `list_selection_view.rs`: side layout geometry, stacked
fallback, buffer clearing, cancel/selection-changed callbacks.
- Integration test in `lib.rs`: theme warning uses the final
(post-resume) config.
## Cargo Deny: Unmaintained Dependency Exceptions
This PR adds two `cargo deny` advisory exceptions for transitive
dependencies pulled in by `syntect v5.3.0`:
| Advisory | Crate | Status |
|----------|-------|--------|
| RUSTSEC-2024-0320 | `yaml-rust` | Unmaintained (maintainer
unreachable) |
| RUSTSEC-2025-0141 | `bincode` | Unmaintained (development ceased;
v1.3.3 considered complete) |
**Why this is safe in our usage:**
- Neither advisory describes a known security vulnerability. Both are
"unmaintained" notices only.
- `bincode` is used by syntect to deserialize pre-compiled syntax sets.
Again, these are **static vendored artifacts** baked into the binary at
build time. No user-supplied bincode data is ever deserialized. - Attack
surface is zero for both crates; exploitation would require a
supply-chain compromise of our own build artifacts.
- These exceptions can be removed when syntect migrates to `yaml-rust2`
and drops `bincode`, or when alternative crates are available upstream.
robust to Nix shell paths (#12476)
## Summary
- Updated `codex-rs/core/src/shell.rs` tests for shell detection to stop
asserting hardcoded shell paths.
- `detects_bash` and `detects_sh` now assert executable basenames
(`bash`, `sh`) rather than `/bin/*`/`/usr/bin/*` absolute paths.
- This keeps behavior the same while avoiding failures in Nix
environments where shells are resolved from `/nix/store/.../bin`.
## Testing
- `nix develop .#default --command sh -lc 'export
PKG_CONFIG_PATH=/nix/store/6az1q591wwlgazzskngr6rl7gmhpyvnc-libcap-2.77-dev/lib/pkgconfig:/nix/store/fgm3pz8486ksh3f94629lpb7xjr2wjp7-openssl-3.6.0-dev/lib/pkgconfig:$PKG_CONFIG_PATH;
export PKG_CONFIG_PATH_FOR_TARGET=$PKG_CONFIG_PATH; cd
/home/alex/workspace/openai/codex/codex-rs && cargo test -p codex-core
--lib detects_bash && cargo test -p codex-core --lib detects_sh'`
## Why
The two failing tests previously hardcoded fixed paths and failed under
the Nix shell due to Nix-provided shell binary locations.
## Links
- Bug report / enhancement request: not publicly filed yet; this was
reproduced in the local Nix environment.
## Why
`codex-core::all` has a flaky test,
`suite::realtime_conversation::conversation_start_audio_text_close_round_trip`,
that assumes a fixed ordering between `conversation.item.create` and
`response.input_audio.delta` requests.
That ordering is not guaranteed: realtime text and audio input are
forwarded through separate queues and a background task, so either
request can be observed first while still being correct behavior.
## What Changed
- Updated the assertion in
`codex-rs/core/tests/suite/realtime_conversation.rs` to compare the two
observed request types order-independently.
- Kept the existing checks that `session.create` is sent first and that
exactly two follow-up requests are recorded.
## Verification
- Re-ran `cargo test -p codex-core --test all
conversation_start_audio_text_close_round_trip` 10 times locally.
Fixes#11845.
Adjust context/token estimation for inline image `data:*;base64,...`
URLs so we
do not count the raw base64 payload as model-visible text.
What changed:
- keep the existing JSON-length estimator as the baseline
- detect only inline base64 `data:` image URLs in message and
function-call
output content items
- subtract only the base64 payload bytes (preserving data URL prefix +
JSON
overhead)
- add a fixed per-image estimate of 340 bytes (~85 tokens at the repo’s
4-bytes/token heuristic)
This avoids large overestimates from MCP image tool outputs while
leaving normal
image URLs (`https://`, `file://`, non-base64 `data:` URLs) unchanged.
Tests:
- message image data URL estimate regression
- function-call output image data URL estimate regression
- non-base64 image URLs unchanged
- non-base64 `data:` URLs unchanged
- `data:application/octet-stream;base64,...` adjusted
- multiple inline images apply multiple fixed costs
- text-only items unchanged
## Why
`codex-core` was carrying the embedded system-skill sample assets (and a
`build.rs` that walks those files to register rerun triggers). Those
assets change infrequently, but any change under `codex-core` still ties
them to `codex-core`'s build/cache lifecycle.
This change moves the embedded system-skills packaging into a dedicated
`codex-skills` crate so it can be cached independently. That reduces
unnecessary invalidation/rebuild pressure on `codex-core` when the
skills bundle is the only thing that changes.
## What Changed
- Added a new `codex-rs/skills` crate (`codex-skills`) with:
- `Cargo.toml`
- `BUILD.bazel`
- `build.rs` to track skill asset file changes for Cargo rebuilds
- `src/lib.rs` containing the embedded system-skills install/cache logic
previously in `codex-core`
- Moved the embedded sample skill assets from
`codex-rs/core/src/skills/assets/samples` to
`codex-rs/skills/src/assets/samples`.
- Updated `codex-rs/core/Cargo.toml` to depend on `codex-skills` and
removed `codex-core`'s direct `include_dir` dependency.
- Removed `codex-core`'s `build.rs`.
- Replaced `codex-rs/core/src/skills/system.rs` implementation with a
thin re-export wrapper to keep existing `codex-core` call sites
unchanged.
- Updated workspace manifests/lockfile (`codex-rs/Cargo.toml`,
`codex-rs/Cargo.lock`) for the new crate.
## Why
`codex-rs/arg0` only needed two things from `codex-core`:
- the `find_codex_home()` wrapper
- the special argv flag used for the internal `apply_patch`
self-invocation path
That made `codex-arg0` depend on `codex-core` for a very small surface
area. This change removes that dependency edge and moves the shared
`apply_patch` invocation flag to a more natural boundary
(`codex-apply-patch`) while keeping the contract explicitly documented.
## What Changed
- Moved the internal `apply_patch` argv[1] flag constant out of
`codex-core` and into `codex-apply-patch`.
- Renamed the constant to `CODEX_CORE_APPLY_PATCH_ARG1` and documented
that it is part of the Codex core process-invocation contract (even
though it now lives in `codex-apply-patch`).
- Updated `arg0`, the core apply-patch runtime, and the `codex-exec`
apply-patch test to import the constant from `codex-apply-patch`.
- Updated `codex-rs/arg0` to call
`codex_utils_home_dir::find_codex_home()` directly instead of
`codex_core::config::find_codex_home()`.
- Removed the `codex-core` dependency from `codex-rs/arg0` and added the
needed direct dependency on `codex-utils-home-dir`.
- Added `codex-apply-patch` as a dev-dependency for `codex-rs/exec`
tests (the apply-patch test now imports the moved constant directly).
## Verification
- `cargo test -p codex-apply-patch`
- `cargo test -p codex-arg0`
- `cargo test -p codex-core --lib apply_patch`
- `cargo test -p codex-exec
test_standalone_exec_cli_can_use_apply_patch`
- `cargo shear`
## Why
`codex-rs/core/src/lib.rs` re-exported a broad set of types and modules
from `codex-protocol` and `codex-shell-command`. That made it easy for
workspace crates to import those APIs through `codex-core`, which in
turn hides dependency edges and makes it harder to reduce compile-time
coupling over time.
This change removes those public re-exports so call sites must import
from the source crates directly. Even when a crate still depends on
`codex-core` today, this makes dependency boundaries explicit and
unblocks future work to drop `codex-core` dependencies where possible.
## What Changed
- Removed public re-exports from `codex-rs/core/src/lib.rs` for:
- `codex_protocol::protocol` and related protocol/model types (including
`InitialHistory`)
- `codex_protocol::config_types` (`protocol_config_types`)
- `codex_shell_command::{bash, is_dangerous_command, is_safe_command,
parse_command, powershell}`
- Migrated workspace Rust call sites to import directly from:
- `codex_protocol::protocol`
- `codex_protocol::config_types`
- `codex_protocol::models`
- `codex_shell_command`
- Added explicit `Cargo.toml` dependencies (`codex-protocol` /
`codex-shell-command`) in crates that now import those crates directly.
- Kept `codex-core` internal modules compiling by using `pub(crate)`
aliases in `core/src/lib.rs` (internal-only, not part of the public
API).
- Updated the two utility crates that can already drop a `codex-core`
dependency edge entirely:
- `codex-utils-approval-presets`
- `codex-utils-cli`
## Verification
- `cargo test -p codex-utils-approval-presets`
- `cargo test -p codex-utils-cli`
- `cargo check --workspace --all-targets`
- `just clippy`
## Why
Compiling `codex-rs/core` is a bottleneck for local iteration, so this
change continues the ongoing extraction of config-related functionality
out of `codex-core` and into `codex-config`.
The goal is not just to move code, but to reduce `codex-core` ownership
and indirection so more code depends on `codex-config` directly.
## What Changed
- Moved config diagnostics logic from
`core/src/config_loader/diagnostics.rs` into
`config/src/diagnostics.rs`.
- Updated `codex-core` to use `codex-config` diagnostics types/functions
directly where possible.
- Removed the `core/src/config_loader/diagnostics.rs` shim module
entirely; the remaining `ConfigToml`-specific calls are in
`core/src/config_loader/mod.rs`.
- Moved `CONFIG_TOML_FILE` into `codex-config` and updated existing
references to use `codex_config::CONFIG_TOML_FILE` directly.
- Added a direct `codex-config` dependency to `codex-cli` for its
`CONFIG_TOML_FILE` use.
## Summary
- move regular-turn context diff/full-context persistence into
`run_turn` so pre-turn compaction runs before incoming context updates
are recorded
- after successful pre-turn compaction, rely on a cleared
`reference_context_item` to trigger full context reinjection on the
follow-up regular turn (manual `/compact` keeps replacement history
summary-only and also clears the baseline)
- preserve `<model_switch>` when full context is reinjected, and inject
it *before* the rest of the full-context items
- scope `reference_context_item` and `previous_model` to regular user
turns only so standalone tasks (`/compact`, shell, review, undo) cannot
suppress future reinjection or `<model_switch>` behavior
- make context-diff persistence + `reference_context_item` updates
explicit in the regular-turn path, with clearer docs/comments around the
invariant
- stop persisting local `/compact` `RolloutItem::TurnContext` snapshots
(only regular turns persist `TurnContextItem` now)
- simplify resume/fork previous-model/reference-baseline hydration by
looking up the last surviving turn context from rollout lifecycle
events, including rollback and compaction-crossing handling
- remove the legacy fallback that guessed from bare `TurnContext`
rollouts without lifecycle events
- update compaction/remote-compaction/model-visible snapshots and
compact test assertions (including remote compaction mock response
shape)
## Why
We were persisting incoming context items before spawning the regular
turn task, which let pre-turn compaction requests accidentally include
incoming context diffs without the new user message. Fixing that exposed
follow-on baseline issues around `/compact`, resume/fork, and standalone
tasks that could cause duplicate context injection or suppress
`<model_switch>` instructions.
This PR re-centers the invariants around regular turns:
- regular turns persist model-visible context diffs/full reinjection and
update the `reference_context_item`
- standalone tasks do not advance those regular-turn baselines
- compaction clears the baseline when replacement history may have
stripped the referenced context diffs
## Follow-ups (TODOs left in code)
- `TODO(ccunningham)`: fix rollback/backtracking baseline handling more
comprehensively
- `TODO(ccunningham)`: include pending incoming context items in
pre-turn compaction threshold estimation
- `TODO(ccunningham)`: inject updated personality spec alongside
`<model_switch>` so some model-switch paths can avoid forced full
reinjection
- `TODO(ccunningham)`: review task turn lifecycle
(`TurnStarted`/`TurnComplete`) behavior and emit task-start context
diffs for task types that should have them (excluding `/compact`)
## Validation
- `just fmt`
- CI should cover the updated compaction/resume/model-visible snapshot
expectations and rollout-hydration behavior
- I did **not** rerun the full local test suite after the latest
resume-lookup / rollout-persistence simplifications
## Summary
`gpt-5.3-codex` really likes to write complicated shell scripts, and
suggest a partial prefix_rule that wouldn't actually approve the
command. We should only show the `prefix_rule` suggestion from the model
if it would actually fully approve the command the user is seeing.
This will technically cause more instances of overly-specific
suggestions when we fallback, but I think the UX is clearer,
particularly when the model doesn't necessarily understand the current
limitations of execpolicy parsing.
## Testing
- [x] Add unit tests
- [x] Add integration tests
- add top-level `experimental_realtime_ws_backend_prompt` config key
(experimental / do not use) and include it in config schema
- apply the override only to `Op::RealtimeConversation` websocket
`backend_prompt`, with config + realtime tests
Michael BolinandGitHub
pakrym-oaiandGitHub
Yaroslav VolovichandGitHub
sayan-oaiandGitHub
Dylan HurdandGitHub
viyatb-oaiandGitHub
Felipe CouryandGitHub
Alex KwiatkowskiandGitHub
Eric TrautandGitHub
Charley CunninghamandGitHub