Merge branch 'main' into codex/add-raw-chain-of-thought-in-codex-cli

codex-rs/core/src/apply_patch.rs

@@ -0,0 +1,434 @@
+use crate::codex::Session;
+use crate::models::FunctionCallOutputPayload;
+use crate::models::ResponseInputItem;
+use crate::protocol::Event;
+use crate::protocol::EventMsg;
+use crate::protocol::FileChange;
+use crate::protocol::PatchApplyBeginEvent;
+use crate::protocol::PatchApplyEndEvent;
+use crate::protocol::ReviewDecision;
+use crate::safety::SafetyCheck;
+use crate::safety::assess_patch_safety;
+use anyhow::Context;
+use codex_apply_patch::AffectedPaths;
+use codex_apply_patch::ApplyPatchAction;
+use codex_apply_patch::ApplyPatchFileChange;
+use codex_apply_patch::print_summary;
+use std::collections::HashMap;
+use std::path::Path;
+use std::path::PathBuf;
+
+pub const CODEX_APPLY_PATCH_ARG1: &str = "--codex-run-as-apply-patch";
+
+pub(crate) enum InternalApplyPatchInvocation {
+    /// The `apply_patch` call was handled programmatically, without any sort
+    /// of sandbox, because the user explicitly approved it. This is the
+    /// result to use with the `shell` function call that contained `apply_patch`.
+    Output(ResponseInputItem),
+
+    /// The `apply_patch` call was auto-approved, which means that, on the
+    /// surface, it appears to be safe, but it should be run in a sandbox if the
+    /// user has configured one because a path being written could be a hard
+    /// link to a file outside the writable folders, so only the sandbox can
+    /// faithfully prevent the write in that case.
+    DelegateToExec(ApplyPatchAction),
+}
+
+impl From<ResponseInputItem> for InternalApplyPatchInvocation {
+    fn from(item: ResponseInputItem) -> Self {
+        InternalApplyPatchInvocation::Output(item)
+    }
+}
+
+pub(crate) async fn apply_patch(
+    sess: &Session,
+    sub_id: &str,
+    call_id: &str,
+    action: ApplyPatchAction,
+) -> InternalApplyPatchInvocation {
+    let writable_roots_snapshot = {
+        #[allow(clippy::unwrap_used)]
+        let guard = sess.writable_roots.lock().unwrap();
+        guard.clone()
+    };
+
+    let auto_approved = match assess_patch_safety(
+        &action,
+        sess.approval_policy,
+        &writable_roots_snapshot,
+        &sess.cwd,
+    ) {
+        SafetyCheck::AutoApprove { .. } => {
+            return InternalApplyPatchInvocation::DelegateToExec(action);
+        }
+        SafetyCheck::AskUser => {
+            // Compute a readable summary of path changes to include in the
+            // approval request so the user can make an informed decision.
+            let rx_approve = sess
+                .request_patch_approval(sub_id.to_owned(), call_id.to_owned(), &action, None, None)
+                .await;
+            match rx_approve.await.unwrap_or_default() {
+                ReviewDecision::Approved | ReviewDecision::ApprovedForSession => false,
+                ReviewDecision::Denied | ReviewDecision::Abort => {
+                    return ResponseInputItem::FunctionCallOutput {
+                        call_id: call_id.to_owned(),
+                        output: FunctionCallOutputPayload {
+                            content: "patch rejected by user".to_string(),
+                            success: Some(false),
+                        },
+                    }
+                    .into();
+                }
+            }
+        }
+        SafetyCheck::Reject { reason } => {
+            return ResponseInputItem::FunctionCallOutput {
+                call_id: call_id.to_owned(),
+                output: FunctionCallOutputPayload {
+                    content: format!("patch rejected: {reason}"),
+                    success: Some(false),
+                },
+            }
+            .into();
+        }
+    };
+
+    // Verify write permissions before touching the filesystem.
+    let writable_snapshot = {
+        #[allow(clippy::unwrap_used)]
+        sess.writable_roots.lock().unwrap().clone()
+    };
+
+    if let Some(offending) = first_offending_path(&action, &writable_snapshot, &sess.cwd) {
+        let root = offending.parent().unwrap_or(&offending).to_path_buf();
+
+        let reason = Some(format!(
+            "grant write access to {} for this session",
+            root.display()
+        ));
+
+        let rx = sess
+            .request_patch_approval(
+                sub_id.to_owned(),
+                call_id.to_owned(),
+                &action,
+                reason.clone(),
+                Some(root.clone()),
+            )
+            .await;
+
+        if !matches!(
+            rx.await.unwrap_or_default(),
+            ReviewDecision::Approved | ReviewDecision::ApprovedForSession
+        ) {
+            return ResponseInputItem::FunctionCallOutput {
+                call_id: call_id.to_owned(),
+                output: FunctionCallOutputPayload {
+                    content: "patch rejected by user".to_string(),
+                    success: Some(false),
+                },
+            }
+            .into();
+        }
+
+        // user approved, extend writable roots for this session
+        #[allow(clippy::unwrap_used)]
+        sess.writable_roots.lock().unwrap().push(root);
+    }
+
+    let _ = sess
+        .tx_event
+        .send(Event {
+            id: sub_id.to_owned(),
+            msg: EventMsg::PatchApplyBegin(PatchApplyBeginEvent {
+                call_id: call_id.to_owned(),
+                auto_approved,
+                changes: convert_apply_patch_to_protocol(&action),
+            }),
+        })
+        .await;
+
+    let mut stdout = Vec::new();
+    let mut stderr = Vec::new();
+    // Enforce writable roots. If a write is blocked, collect offending root
+    // and prompt the user to extend permissions.
+    let mut result = apply_changes_from_apply_patch_and_report(&action, &mut stdout, &mut stderr);
+
+    if let Err(err) = &result {
+        if err.kind() == std::io::ErrorKind::PermissionDenied {
+            // Determine first offending path.
+            let offending_opt = action
+                .changes()
+                .iter()
+                .flat_map(|(path, change)| match change {
+                    ApplyPatchFileChange::Add { .. } => vec![path.as_ref()],
+                    ApplyPatchFileChange::Delete => vec![path.as_ref()],
+                    ApplyPatchFileChange::Update {
+                        move_path: Some(move_path),
+                        ..
+                    } => {
+                        vec![path.as_ref(), move_path.as_ref()]
+                    }
+                    ApplyPatchFileChange::Update {
+                        move_path: None, ..
+                    } => vec![path.as_ref()],
+                })
+                .find_map(|path: &Path| {
+                    // ApplyPatchAction promises to guarantee absolute paths.
+                    if !path.is_absolute() {
+                        panic!("apply_patch invariant failed: path is not absolute: {path:?}");
+                    }
+
+                    let writable = {
+                        #[allow(clippy::unwrap_used)]
+                        let roots = sess.writable_roots.lock().unwrap();
+                        roots.iter().any(|root| path.starts_with(root))
+                    };
+                    if writable {
+                        None
+                    } else {
+                        Some(path.to_path_buf())
+                    }
+                });
+
+            if let Some(offending) = offending_opt {
+                let root = offending.parent().unwrap_or(&offending).to_path_buf();
+
+                let reason = Some(format!(
+                    "grant write access to {} for this session",
+                    root.display()
+                ));
+                let rx = sess
+                    .request_patch_approval(
+                        sub_id.to_owned(),
+                        call_id.to_owned(),
+                        &action,
+                        reason.clone(),
+                        Some(root.clone()),
+                    )
+                    .await;
+                if matches!(
+                    rx.await.unwrap_or_default(),
+                    ReviewDecision::Approved | ReviewDecision::ApprovedForSession
+                ) {
+                    // Extend writable roots.
+                    #[allow(clippy::unwrap_used)]
+                    sess.writable_roots.lock().unwrap().push(root);
+                    stdout.clear();
+                    stderr.clear();
+                    result = apply_changes_from_apply_patch_and_report(
+                        &action,
+                        &mut stdout,
+                        &mut stderr,
+                    );
+                }
+            }
+        }
+    }
+
+    // Emit PatchApplyEnd event.
+    let success_flag = result.is_ok();
+    let _ = sess
+        .tx_event
+        .send(Event {
+            id: sub_id.to_owned(),
+            msg: EventMsg::PatchApplyEnd(PatchApplyEndEvent {
+                call_id: call_id.to_owned(),
+                stdout: String::from_utf8_lossy(&stdout).to_string(),
+                stderr: String::from_utf8_lossy(&stderr).to_string(),
+                success: success_flag,
+            }),
+        })
+        .await;
+
+    let item = match result {
+        Ok(_) => ResponseInputItem::FunctionCallOutput {
+            call_id: call_id.to_owned(),
+            output: FunctionCallOutputPayload {
+                content: String::from_utf8_lossy(&stdout).to_string(),
+                success: None,
+            },
+        },
+        Err(e) => ResponseInputItem::FunctionCallOutput {
+            call_id: call_id.to_owned(),
+            output: FunctionCallOutputPayload {
+                content: format!("error: {e:#}, stderr: {}", String::from_utf8_lossy(&stderr)),
+                success: Some(false),
+            },
+        },
+    };
+    InternalApplyPatchInvocation::Output(item)
+}
+
+/// Return the first path in `hunks` that is NOT under any of the
+/// `writable_roots` (after normalising). If all paths are acceptable,
+/// returns None.
+fn first_offending_path(
+    action: &ApplyPatchAction,
+    writable_roots: &[PathBuf],
+    cwd: &Path,
+) -> Option<PathBuf> {
+    let changes = action.changes();
+    for (path, change) in changes {
+        let candidate = match change {
+            ApplyPatchFileChange::Add { .. } => path,
+            ApplyPatchFileChange::Delete => path,
+            ApplyPatchFileChange::Update { move_path, .. } => move_path.as_ref().unwrap_or(path),
+        };
+
+        let abs = if candidate.is_absolute() {
+            candidate.clone()
+        } else {
+            cwd.join(candidate)
+        };
+
+        let mut allowed = false;
+        for root in writable_roots {
+            let root_abs = if root.is_absolute() {
+                root.clone()
+            } else {
+                cwd.join(root)
+            };
+            if abs.starts_with(&root_abs) {
+                allowed = true;
+                break;
+            }
+        }
+
+        if !allowed {
+            return Some(candidate.clone());
+        }
+    }
+    None
+}
+
+pub(crate) fn convert_apply_patch_to_protocol(
+    action: &ApplyPatchAction,
+) -> HashMap<PathBuf, FileChange> {
+    let changes = action.changes();
+    let mut result = HashMap::with_capacity(changes.len());
+    for (path, change) in changes {
+        let protocol_change = match change {
+            ApplyPatchFileChange::Add { content } => FileChange::Add {
+                content: content.clone(),
+            },
+            ApplyPatchFileChange::Delete => FileChange::Delete,
+            ApplyPatchFileChange::Update {
+                unified_diff,
+                move_path,
+                new_content: _new_content,
+            } => FileChange::Update {
+                unified_diff: unified_diff.clone(),
+                move_path: move_path.clone(),
+            },
+        };
+        result.insert(path.clone(), protocol_change);
+    }
+    result
+}
+
+fn apply_changes_from_apply_patch_and_report(
+    action: &ApplyPatchAction,
+    stdout: &mut impl std::io::Write,
+    stderr: &mut impl std::io::Write,
+) -> std::io::Result<()> {
+    match apply_changes_from_apply_patch(action) {
+        Ok(affected_paths) => {
+            print_summary(&affected_paths, stdout)?;
+        }
+        Err(err) => {
+            writeln!(stderr, "{err:?}")?;
+        }
+    }
+
+    Ok(())
+}
+
+fn apply_changes_from_apply_patch(action: &ApplyPatchAction) -> anyhow::Result<AffectedPaths> {
+    let mut added: Vec<PathBuf> = Vec::new();
+    let mut modified: Vec<PathBuf> = Vec::new();
+    let mut deleted: Vec<PathBuf> = Vec::new();
+
+    let changes = action.changes();
+    for (path, change) in changes {
+        match change {
+            ApplyPatchFileChange::Add { content } => {
+                if let Some(parent) = path.parent() {
+                    if !parent.as_os_str().is_empty() {
+                        std::fs::create_dir_all(parent).with_context(|| {
+                            format!("Failed to create parent directories for {}", path.display())
+                        })?;
+                    }
+                }
+                std::fs::write(path, content)
+                    .with_context(|| format!("Failed to write file {}", path.display()))?;
+                added.push(path.clone());
+            }
+            ApplyPatchFileChange::Delete => {
+                std::fs::remove_file(path)
+                    .with_context(|| format!("Failed to delete file {}", path.display()))?;
+                deleted.push(path.clone());
+            }
+            ApplyPatchFileChange::Update {
+                unified_diff: _unified_diff,
+                move_path,
+                new_content,
+            } => {
+                if let Some(move_path) = move_path {
+                    if let Some(parent) = move_path.parent() {
+                        if !parent.as_os_str().is_empty() {
+                            std::fs::create_dir_all(parent).with_context(|| {
+                                format!(
+                                    "Failed to create parent directories for {}",
+                                    move_path.display()
+                                )
+                            })?;
+                        }
+                    }
+
+                    std::fs::rename(path, move_path)
+                        .with_context(|| format!("Failed to rename file {}", path.display()))?;
+                    std::fs::write(move_path, new_content)?;
+                    modified.push(move_path.clone());
+                    deleted.push(path.clone());
+                } else {
+                    std::fs::write(path, new_content)?;
+                    modified.push(path.clone());
+                }
+            }
+        }
+    }
+
+    Ok(AffectedPaths {
+        added,
+        modified,
+        deleted,
+    })
+}
+
+pub(crate) fn get_writable_roots(cwd: &Path) -> Vec<PathBuf> {
+    let mut writable_roots = Vec::new();
+    if cfg!(target_os = "macos") {
+        // On macOS, $TMPDIR is private to the user.
+        writable_roots.push(std::env::temp_dir());
+
+        // Allow pyenv to update its shims directory. Without this, any tool
+        // that happens to be managed by `pyenv` will fail with an error like:
+        //
+        //   pyenv: cannot rehash: $HOME/.pyenv/shims isn't writable
+        //
+        // which is emitted every time `pyenv` tries to run `rehash` (for
+        // example, after installing a new Python package that drops an entry
+        // point). Although the sandbox is intentionally read‑only by default,
+        // writing to the user's local `pyenv` directory is safe because it
+        // is already user‑writable and scoped to the current user account.
+        if let Ok(home_dir) = std::env::var("HOME") {
+            let pyenv_dir = PathBuf::from(home_dir).join(".pyenv");
+            writable_roots.push(pyenv_dir);
+        }
+    }
+
+    writable_roots.push(cwd.to_path_buf());
+
+    writable_roots
+}

codex-rs/core/src/bash.rs

@@ -0,0 +1,219 @@
+use tree_sitter::Parser;
+use tree_sitter::Tree;
+use tree_sitter_bash::LANGUAGE as BASH;
+
+/// Parse the provided bash source using tree-sitter-bash, returning a Tree on
+/// success or None if parsing failed.
+pub fn try_parse_bash(bash_lc_arg: &str) -> Option<Tree> {
+    let lang = BASH.into();
+    let mut parser = Parser::new();
+    #[expect(clippy::expect_used)]
+    parser.set_language(&lang).expect("load bash grammar");
+    let old_tree: Option<&Tree> = None;
+    parser.parse(bash_lc_arg, old_tree)
+}
+
+/// Parse a script which may contain multiple simple commands joined only by
+/// the safe logical/pipe/sequencing operators: `&&`, `||`, `;`, `|`.
+///
+/// Returns `Some(Vec<command_words>)` if every command is a plain word‑only
+/// command and the parse tree does not contain disallowed constructs
+/// (parentheses, redirections, substitutions, control flow, etc.). Otherwise
+/// returns `None`.
+pub fn try_parse_word_only_commands_sequence(tree: &Tree, src: &str) -> Option<Vec<Vec<String>>> {
+    if tree.root_node().has_error() {
+        return None;
+    }
+
+    // List of allowed (named) node kinds for a "word only commands sequence".
+    // If we encounter a named node that is not in this list we reject.
+    const ALLOWED_KINDS: &[&str] = &[
+        // top level containers
+        "program",
+        "list",
+        "pipeline",
+        // commands & words
+        "command",
+        "command_name",
+        "word",
+        "string",
+        "string_content",
+        "raw_string",
+        "number",
+    ];
+    // Allow only safe punctuation / operator tokens; anything else causes reject.
+    const ALLOWED_PUNCT_TOKENS: &[&str] = &["&&", "||", ";", "|", "\"", "'"];
+
+    let root = tree.root_node();
+    let mut cursor = root.walk();
+    let mut stack = vec![root];
+    let mut command_nodes = Vec::new();
+    while let Some(node) = stack.pop() {
+        let kind = node.kind();
+        if node.is_named() {
+            if !ALLOWED_KINDS.contains(&kind) {
+                return None;
+            }
+            if kind == "command" {
+                command_nodes.push(node);
+            }
+        } else {
+            // Reject any punctuation / operator tokens that are not explicitly allowed.
+            if kind.chars().any(|c| "&;|".contains(c)) && !ALLOWED_PUNCT_TOKENS.contains(&kind) {
+                return None;
+            }
+            if !(ALLOWED_PUNCT_TOKENS.contains(&kind) || kind.trim().is_empty()) {
+                // If it's a quote token or operator it's allowed above; we also allow whitespace tokens.
+                // Any other punctuation like parentheses, braces, redirects, backticks, etc are rejected.
+                return None;
+            }
+        }
+        for child in node.children(&mut cursor) {
+            stack.push(child);
+        }
+    }
+
+    let mut commands = Vec::new();
+    for node in command_nodes {
+        if let Some(words) = parse_plain_command_from_node(node, src) {
+            commands.push(words);
+        } else {
+            return None;
+        }
+    }
+    Some(commands)
+}
+
+fn parse_plain_command_from_node(cmd: tree_sitter::Node, src: &str) -> Option<Vec<String>> {
+    if cmd.kind() != "command" {
+        return None;
+    }
+    let mut words = Vec::new();
+    let mut cursor = cmd.walk();
+    for child in cmd.named_children(&mut cursor) {
+        match child.kind() {
+            "command_name" => {
+                let word_node = child.named_child(0)?;
+                if word_node.kind() != "word" {
+                    return None;
+                }
+                words.push(word_node.utf8_text(src.as_bytes()).ok()?.to_owned());
+            }
+            "word" | "number" => {
+                words.push(child.utf8_text(src.as_bytes()).ok()?.to_owned());
+            }
+            "string" => {
+                if child.child_count() == 3
+                    && child.child(0)?.kind() == "\""
+                    && child.child(1)?.kind() == "string_content"
+                    && child.child(2)?.kind() == "\""
+                {
+                    words.push(child.child(1)?.utf8_text(src.as_bytes()).ok()?.to_owned());
+                } else {
+                    return None;
+                }
+            }
+            "raw_string" => {
+                let raw_string = child.utf8_text(src.as_bytes()).ok()?;
+                let stripped = raw_string
+                    .strip_prefix('\'')
+                    .and_then(|s| s.strip_suffix('\''));
+                if let Some(s) = stripped {
+                    words.push(s.to_owned());
+                } else {
+                    return None;
+                }
+            }
+            _ => return None,
+        }
+    }
+    Some(words)
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::unwrap_used)]
+    use super::*;
+
+    fn parse_seq(src: &str) -> Option<Vec<Vec<String>>> {
+        let tree = try_parse_bash(src)?;
+        try_parse_word_only_commands_sequence(&tree, src)
+    }
+
+    #[test]
+    fn accepts_single_simple_command() {
+        let cmds = parse_seq("ls -1").unwrap();
+        assert_eq!(cmds, vec![vec!["ls".to_string(), "-1".to_string()]]);
+    }
+
+    #[test]
+    fn accepts_multiple_commands_with_allowed_operators() {
+        let src = "ls && pwd; echo 'hi there' | wc -l";
+        let cmds = parse_seq(src).unwrap();
+        let expected: Vec<Vec<String>> = vec![
+            vec!["wc".to_string(), "-l".to_string()],
+            vec!["echo".to_string(), "hi there".to_string()],
+            vec!["pwd".to_string()],
+            vec!["ls".to_string()],
+        ];
+        assert_eq!(cmds, expected);
+    }
+
+    #[test]
+    fn extracts_double_and_single_quoted_strings() {
+        let cmds = parse_seq("echo \"hello world\"").unwrap();
+        assert_eq!(
+            cmds,
+            vec![vec!["echo".to_string(), "hello world".to_string()]]
+        );
+
+        let cmds2 = parse_seq("echo 'hi there'").unwrap();
+        assert_eq!(
+            cmds2,
+            vec![vec!["echo".to_string(), "hi there".to_string()]]
+        );
+    }
+
+    #[test]
+    fn accepts_numbers_as_words() {
+        let cmds = parse_seq("echo 123 456").unwrap();
+        assert_eq!(
+            cmds,
+            vec![vec![
+                "echo".to_string(),
+                "123".to_string(),
+                "456".to_string()
+            ]]
+        );
+    }
+
+    #[test]
+    fn rejects_parentheses_and_subshells() {
+        assert!(parse_seq("(ls)").is_none());
+        assert!(parse_seq("ls || (pwd && echo hi)").is_none());
+    }
+
+    #[test]
+    fn rejects_redirections_and_unsupported_operators() {
+        assert!(parse_seq("ls > out.txt").is_none());
+        assert!(parse_seq("echo hi & echo bye").is_none());
+    }
+
+    #[test]
+    fn rejects_command_and_process_substitutions_and_expansions() {
+        assert!(parse_seq("echo $(pwd)").is_none());
+        assert!(parse_seq("echo `pwd`").is_none());
+        assert!(parse_seq("echo $HOME").is_none());
+        assert!(parse_seq("echo \"hi $USER\"").is_none());
+    }
+
+    #[test]
+    fn rejects_variable_assignment_prefix() {
+        assert!(parse_seq("FOO=bar ls").is_none());
+    }
+
+    #[test]
+    fn rejects_trailing_operator_parse_error() {
+        assert!(parse_seq("ls &&").is_none());
+    }
+}

codex-rs/core/src/chat_completions.rs

@@ -21,8 +21,6 @@ use crate::client_common::ResponseEvent;
 use crate::client_common::ResponseStream;
 use crate::error::CodexErr;
 use crate::error::Result;
-use crate::flags::OPENAI_REQUEST_MAX_RETRIES;
-use crate::flags::OPENAI_STREAM_IDLE_TIMEOUT_MS;
 use crate::models::ContentItem;
 use crate::models::ResponseItem;
 use crate::openai_tools::create_tools_json_for_chat_completions_api;
@@ -32,6 +30,7 @@ use crate::util::backoff;
 pub(crate) async fn stream_chat_completions(
     prompt: &Prompt,
     model: &str,
+    include_plan_tool: bool,
     client: &reqwest::Client,
     provider: &ModelProviderInfo,
 ) -> Result<ResponseStream> {
@@ -41,9 +40,13 @@ pub(crate) async fn stream_chat_completions(
     let full_instructions = prompt.get_full_instructions(model);
     messages.push(json!({"role": "system", "content": full_instructions}));
 
+    if let Some(instr) = &prompt.user_instructions {
+        messages.push(json!({"role": "user", "content": instr}));
+    }
+
     for item in &prompt.input {
         match item {
-            ResponseItem::Message { role, content } => {
+            ResponseItem::Message { role, content, .. } => {
                 let mut text = String::new();
                 for c in content {
                     match c {
@@ -60,6 +63,7 @@ pub(crate) async fn stream_chat_completions(
                 name,
                 arguments,
                 call_id,
+                ..
             } => {
                 messages.push(json!({
                     "role": "assistant",
@@ -106,7 +110,7 @@ pub(crate) async fn stream_chat_completions(
         }
     }
 
-    let tools_json = create_tools_json_for_chat_completions_api(prompt, model)?;
+    let tools_json = create_tools_json_for_chat_completions_api(prompt, model, include_plan_tool)?;
     let payload = json!({
         "model": model,
         "messages": messages,
@@ -121,6 +125,7 @@ pub(crate) async fn stream_chat_completions(
     );
 
     let mut attempt = 0;
+    let max_retries = provider.request_max_retries();
     loop {
         attempt += 1;
 
@@ -134,9 +139,13 @@ pub(crate) async fn stream_chat_completions(
 
         match res {
             Ok(resp) if resp.status().is_success() => {
-                let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(16);
+                let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
                 let stream = resp.bytes_stream().map_err(CodexErr::Reqwest);
-                tokio::spawn(process_chat_sse(stream, tx_event));
+                tokio::spawn(process_chat_sse(
+                    stream,
+                    tx_event,
+                    provider.stream_idle_timeout(),
+                ));
                 return Ok(ResponseStream { rx_event });
             }
             Ok(res) => {
@@ -146,7 +155,7 @@ pub(crate) async fn stream_chat_completions(
                     return Err(CodexErr::UnexpectedStatus(status, body));
                 }
 
-                if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                if attempt > max_retries {
                     return Err(CodexErr::RetryLimit(status));
                 }
 
@@ -162,7 +171,7 @@ pub(crate) async fn stream_chat_completions(
                 tokio::time::sleep(delay).await;
             }
             Err(e) => {
-                if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                if attempt > max_retries {
                     return Err(e.into());
                 }
                 let delay = backoff(attempt);
@@ -175,14 +184,15 @@ pub(crate) async fn stream_chat_completions(
 /// Lightweight SSE processor for the Chat Completions streaming format. The
 /// output is mapped onto Codex's internal [`ResponseEvent`] so that the rest
 /// of the pipeline can stay agnostic of the underlying wire format.
-async fn process_chat_sse<S>(stream: S, tx_event: mpsc::Sender<Result<ResponseEvent>>)
-where
+async fn process_chat_sse<S>(
+    stream: S,
+    tx_event: mpsc::Sender<Result<ResponseEvent>>,
+    idle_timeout: Duration,
+) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
     let mut stream = stream.eventsource();
 
-    let idle_timeout = *OPENAI_STREAM_IDLE_TIMEOUT_MS;
-
     // State to accumulate a function call across streaming chunks.
     // OpenAI may split the `arguments` string over multiple `delta` events
     // until the chunk whose `finish_reason` is `tool_calls` is emitted. We
@@ -255,6 +265,7 @@ where
                     content: vec![ContentItem::OutputText {
                         text: content.to_string(),
                     }],
+                    id: None,
                 };
 
                 let _ = tx_event.send(Ok(ResponseEvent::OutputItemDone(item))).await;
@@ -296,6 +307,7 @@ where
                     "tool_calls" if fn_call_state.active => {
                         // Build the FunctionCall response item.
                         let item = ResponseItem::FunctionCall {
+                            id: None,
                             name: fn_call_state.name.clone().unwrap_or_else(|| "".to_string()),
                             arguments: fn_call_state.arguments.clone(),
                             call_id: fn_call_state.call_id.clone().unwrap_or_else(String::new),
@@ -398,6 +410,7 @@ where
                 }))) => {
                     if !this.cumulative.is_empty() {
                         let aggregated_item = crate::models::ResponseItem::Message {
+                            id: None,
                             role: "assistant".to_string(),
                             content: vec![crate::models::ContentItem::OutputText {
                                 text: std::mem::take(&mut this.cumulative),
@@ -426,6 +439,12 @@ where
                     // will never appear in a Chat Completions stream.
                     continue;
                 }
+                Poll::Ready(Some(Ok(ResponseEvent::OutputTextDelta(_))))
+                | Poll::Ready(Some(Ok(ResponseEvent::ReasoningSummaryDelta(_)))) => {
+                    // Deltas are ignored here since aggregation waits for the
+                    // final OutputItemDone.
+                    continue;
+                }
             }
         }
     }

codex-rs/core/src/client.rs

@@ -3,6 +3,8 @@ use std::path::Path;
 use std::time::Duration;
 
 use bytes::Bytes;
+use codex_login::AuthMode;
+use codex_login::CodexAuth;
 use eventsource_stream::Eventsource;
 use futures::prelude::*;
 use reqwest::StatusCode;
@@ -15,6 +17,7 @@ use tokio_util::io::ReaderStream;
 use tracing::debug;
 use tracing::trace;
 use tracing::warn;
+use uuid::Uuid;
 
 use crate::chat_completions::AggregateStreamExt;
 use crate::chat_completions::stream_chat_completions;
@@ -27,12 +30,12 @@ use crate::config::Config;
 use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
 use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::error::CodexErr;
+use crate::error::EnvVarError;
 use crate::error::Result;
 use crate::flags::CODEX_RS_SSE_FIXTURE;
-use crate::flags::OPENAI_REQUEST_MAX_RETRIES;
-use crate::flags::OPENAI_STREAM_IDLE_TIMEOUT_MS;
 use crate::model_provider_info::ModelProviderInfo;
 use crate::model_provider_info::WireApi;
+use crate::models::ContentItem;
 use crate::models::ResponseItem;
 use crate::openai_tools::create_tools_json_for_responses_api;
 use crate::protocol::TokenUsage;
@@ -42,8 +45,10 @@ use std::sync::Arc;
 #[derive(Clone)]
 pub struct ModelClient {
     config: Arc<Config>,
+    auth: Option<CodexAuth>,
     client: reqwest::Client,
     provider: ModelProviderInfo,
+    session_id: Uuid,
     effort: ReasoningEffortConfig,
     summary: ReasoningSummaryConfig,
 }
@@ -51,14 +56,18 @@ pub struct ModelClient {
 impl ModelClient {
     pub fn new(
         config: Arc<Config>,
+        auth: Option<CodexAuth>,
         provider: ModelProviderInfo,
         effort: ReasoningEffortConfig,
         summary: ReasoningSummaryConfig,
+        session_id: Uuid,
     ) -> Self {
         Self {
             config,
+            auth,
             client: reqwest::Client::new(),
             provider,
+            session_id,
             effort,
             summary,
         }
@@ -75,6 +84,7 @@ impl ModelClient {
                 let response_stream = stream_chat_completions(
                     prompt,
                     &self.config.model,
+                    self.config.include_plan_tool,
                     &self.client,
                     &self.provider,
                 )
@@ -109,23 +119,65 @@ impl ModelClient {
         if let Some(path) = &*CODEX_RS_SSE_FIXTURE {
             // short circuit for tests
             warn!(path, "Streaming from fixture");
-            return stream_from_fixture(path).await;
+            return stream_from_fixture(path, self.provider.clone()).await;
         }
 
+        let auth = self.auth.as_ref().ok_or_else(|| {
+            CodexErr::EnvVar(EnvVarError {
+                var: "OPENAI_API_KEY".to_string(),
+                instructions: Some("Create an API key (https://platform.openai.com) and export it as an environment variable.".to_string()),
+            })
+        })?;
+
+        let store = prompt.store && auth.mode != AuthMode::ChatGPT;
+
+        let base_url = match self.provider.base_url.clone() {
+            Some(url) => url,
+            None => match auth.mode {
+                AuthMode::ChatGPT => "https://chatgpt.com/backend-api/codex".to_string(),
+                AuthMode::ApiKey => "https://api.openai.com/v1".to_string(),
+            },
+        };
+
+        let token = auth.get_token().await?;
+
         let full_instructions = prompt.get_full_instructions(&self.config.model);
-        let tools_json = create_tools_json_for_responses_api(prompt, &self.config.model)?;
+        let tools_json = create_tools_json_for_responses_api(
+            prompt,
+            &self.config.model,
+            self.config.include_plan_tool,
+        )?;
         let reasoning = create_reasoning_param_for_request(&self.config, self.effort, self.summary);
+
+        // Request encrypted COT if we are not storing responses,
+        // otherwise reasoning items will be referenced by ID
+        let include: Vec<String> = if !store && reasoning.is_some() {
+            vec!["reasoning.encrypted_content".to_string()]
+        } else {
+            vec![]
+        };
+
+        let mut input_with_instructions = Vec::with_capacity(prompt.input.len() + 1);
+        if let Some(ui) = &prompt.user_instructions {
+            input_with_instructions.push(ResponseItem::Message {
+                id: None,
+                role: "user".to_string(),
+                content: vec![ContentItem::InputText { text: ui.clone() }],
+            });
+        }
+        input_with_instructions.extend(prompt.input.clone());
+
         let payload = ResponsesApiRequest {
             model: &self.config.model,
             instructions: &full_instructions,
-            input: &prompt.input,
+            input: &input_with_instructions,
             tools: &tools_json,
             tool_choice: "auto",
             parallel_tool_calls: false,
             reasoning,
-            previous_response_id: prompt.prev_id.clone(),
-            store: prompt.store,
+            store,
             stream: true,
+            include,
         };
 
         trace!(
@@ -135,24 +187,45 @@ impl ModelClient {
         );
 
         let mut attempt = 0;
+        let max_retries = self.provider.request_max_retries();
+
         loop {
             attempt += 1;
 
             let req_builder = self
-                .provider
-                .create_request_builder(&self.client)?
+                .client
+                .post(format!("{base_url}/responses"))
                 .header("OpenAI-Beta", "responses=experimental")
+                .header("session_id", self.session_id.to_string())
+                .bearer_auth(&token)
                 .header(reqwest::header::ACCEPT, "text/event-stream")
                 .json(&payload);
 
+            let req_builder = self.provider.apply_http_headers(req_builder);
+
             let res = req_builder.send().await;
+            if let Ok(resp) = &res {
+                trace!(
+                    "Response status: {}, request-id: {}",
+                    resp.status(),
+                    resp.headers()
+                        .get("x-request-id")
+                        .map(|v| v.to_str().unwrap_or_default())
+                        .unwrap_or_default()
+                );
+            }
+
             match res {
                 Ok(resp) if resp.status().is_success() => {
-                    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(16);
+                    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
 
                     // spawn task to process SSE
                     let stream = resp.bytes_stream().map_err(CodexErr::Reqwest);
-                    tokio::spawn(process_sse(stream, tx_event));
+                    tokio::spawn(process_sse(
+                        stream,
+                        tx_event,
+                        self.provider.stream_idle_timeout(),
+                    ));
 
                     return Ok(ResponseStream { rx_event });
                 }
@@ -171,7 +244,7 @@ impl ModelClient {
                         return Err(CodexErr::UnexpectedStatus(status, body));
                     }
 
-                    if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                    if attempt > max_retries {
                         return Err(CodexErr::RetryLimit(status));
                     }
 
@@ -188,7 +261,7 @@ impl ModelClient {
                     tokio::time::sleep(delay).await;
                 }
                 Err(e) => {
-                    if attempt > *OPENAI_REQUEST_MAX_RETRIES {
+                    if attempt > max_retries {
                         return Err(e.into());
                     }
                     let delay = backoff(attempt);
@@ -197,6 +270,10 @@ impl ModelClient {
             }
         }
     }
+
+    pub fn get_provider(&self) -> ModelProviderInfo {
+        self.provider.clone()
+    }
 }
 
 #[derive(Debug, Deserialize, Serialize)]
@@ -205,6 +282,7 @@ struct SseEvent {
     kind: String,
     response: Option<Value>,
     item: Option<Value>,
+    delta: Option<String>,
 }
 
 #[derive(Debug, Deserialize)]
@@ -247,14 +325,16 @@ struct ResponseCompletedOutputTokensDetails {
     reasoning_tokens: u64,
 }
 
-async fn process_sse<S>(stream: S, tx_event: mpsc::Sender<Result<ResponseEvent>>)
-where
+async fn process_sse<S>(
+    stream: S,
+    tx_event: mpsc::Sender<Result<ResponseEvent>>,
+    idle_timeout: Duration,
+) where
     S: Stream<Item = Result<Bytes>> + Unpin,
 {
     let mut stream = stream.eventsource();
 
     // If the stream stays completely silent for an extended period treat it as disconnected.
-    let idle_timeout = *OPENAI_STREAM_IDLE_TIMEOUT_MS;
     // The response id returned from the "complete" message.
     let mut response_completed: Option<ResponseCompleted> = None;
 
@@ -315,7 +395,7 @@ where
             // duplicated `output` array embedded in the `response.completed`
             // payload.  That produced two concrete issues:
             //   1. No real‑time streaming – the user only saw output after the
-            //      entire turn had finished, which broke the “typing” UX and
+            //      entire turn had finished, which broke the "typing" UX and
             //      made long‑running turns look stalled.
             //   2. Duplicate `function_call_output` items – both the
             //      individual *and* the completed array were forwarded, which
@@ -337,11 +417,40 @@ where
                     return;
                 }
             }
+            "response.output_text.delta" => {
+                if let Some(delta) = event.delta {
+                    let event = ResponseEvent::OutputTextDelta(delta);
+                    if tx_event.send(Ok(event)).await.is_err() {
+                        return;
+                    }
+                }
+            }
+            "response.reasoning_summary_text.delta" => {
+                if let Some(delta) = event.delta {
+                    let event = ResponseEvent::ReasoningSummaryDelta(delta);
+                    if tx_event.send(Ok(event)).await.is_err() {
+                        return;
+                    }
+                }
+            }
             "response.created" => {
                 if event.response.is_some() {
                     let _ = tx_event.send(Ok(ResponseEvent::Created {})).await;
                 }
             }
+            "response.failed" => {
+                if let Some(resp_val) = event.response {
+                    let error = resp_val
+                        .get("error")
+                        .and_then(|v| v.get("message"))
+                        .and_then(|v| v.as_str())
+                        .unwrap_or("response.failed event received");
+
+                    let _ = tx_event
+                        .send(Err(CodexErr::Stream(error.to_string())))
+                        .await;
+                }
+            }
             // Final response completed – includes array of output items & id
             "response.completed" => {
                 if let Some(resp_val) = event.response {
@@ -360,10 +469,8 @@ where
             | "response.function_call_arguments.delta"
             | "response.in_progress"
             | "response.output_item.added"
-            | "response.output_text.delta"
             | "response.output_text.done"
             | "response.reasoning_summary_part.added"
-            | "response.reasoning_summary_text.delta"
             | "response.reasoning_summary_text.done" => {
                 // Currently, we ignore these events, but we handle them
                 // separately to skip the logging message in the `other` case.
@@ -374,8 +481,11 @@ where
 }
 
 /// used in tests to stream from a text SSE file
-async fn stream_from_fixture(path: impl AsRef<Path>) -> Result<ResponseStream> {
-    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(16);
+async fn stream_from_fixture(
+    path: impl AsRef<Path>,
+    provider: ModelProviderInfo,
+) -> Result<ResponseStream> {
+    let (tx_event, rx_event) = mpsc::channel::<Result<ResponseEvent>>(1600);
     let f = std::fs::File::open(path.as_ref())?;
     let lines = std::io::BufReader::new(f).lines();
 
@@ -388,17 +498,57 @@ async fn stream_from_fixture(path: impl AsRef<Path>) -> Result<ResponseStream> {
 
     let rdr = std::io::Cursor::new(content);
     let stream = ReaderStream::new(rdr).map_err(CodexErr::Io);
-    tokio::spawn(process_sse(stream, tx_event));
+    tokio::spawn(process_sse(
+        stream,
+        tx_event,
+        provider.stream_idle_timeout(),
+    ));
     Ok(ResponseStream { rx_event })
 }
 
 #[cfg(test)]
 mod tests {
     #![allow(clippy::expect_used, clippy::unwrap_used)]
+
     use super::*;
     use serde_json::json;
+    use tokio::sync::mpsc;
+    use tokio_test::io::Builder as IoBuilder;
+    use tokio_util::io::ReaderStream;
 
-    async fn run_sse(events: Vec<serde_json::Value>) -> Vec<ResponseEvent> {
+    // ────────────────────────────
+    // Helpers
+    // ────────────────────────────
+
+    /// Runs the SSE parser on pre-chunked byte slices and returns every event
+    /// (including any final `Err` from a stream-closure check).
+    async fn collect_events(
+        chunks: &[&[u8]],
+        provider: ModelProviderInfo,
+    ) -> Vec<Result<ResponseEvent>> {
+        let mut builder = IoBuilder::new();
+        for chunk in chunks {
+            builder.read(chunk);
+        }
+
+        let reader = builder.build();
+        let stream = ReaderStream::new(reader).map_err(CodexErr::Io);
+        let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(16);
+        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));
+
+        let mut events = Vec::new();
+        while let Some(ev) = rx.recv().await {
+            events.push(ev);
+        }
+        events
+    }
+
+    /// Builds an in-memory SSE stream from JSON fixtures and returns only the
+    /// successfully parsed events (panics on internal channel errors).
+    async fn run_sse(
+        events: Vec<serde_json::Value>,
+        provider: ModelProviderInfo,
+    ) -> Vec<ResponseEvent> {
         let mut body = String::new();
         for e in events {
             let kind = e
@@ -411,9 +561,11 @@ mod tests {
                 body.push_str(&format!("event: {kind}\ndata: {e}\n\n"));
             }
         }
+
         let (tx, mut rx) = mpsc::channel::<Result<ResponseEvent>>(8);
         let stream = ReaderStream::new(std::io::Cursor::new(body)).map_err(CodexErr::Io);
-        tokio::spawn(process_sse(stream, tx));
+        tokio::spawn(process_sse(stream, tx, provider.stream_idle_timeout()));
+
         let mut out = Vec::new();
         while let Some(ev) = rx.recv().await {
             out.push(ev.expect("channel closed"));
@@ -421,14 +573,137 @@ mod tests {
         out
     }
 
-    /// Verifies that the SSE adapter emits the expected [`ResponseEvent`] for
-    /// a variety of `type` values from the Responses API. The test is written
-    /// table-driven style to keep additions for new event kinds trivial.
-    ///
-    /// Each `Case` supplies an input event, a predicate that must match the
-    /// *first* `ResponseEvent` produced by the adapter, and the total number
-    /// of events expected after appending a synthetic `response.completed`
-    /// marker that terminates the stream.
+    // ────────────────────────────
+    // Tests from `implement-test-for-responses-api-sse-parser`
+    // ────────────────────────────
+
+    #[tokio::test]
+    async fn parses_items_and_completed() {
+        let item1 = json!({
+            "type": "response.output_item.done",
+            "item": {
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "Hello"}]
+            }
+        })
+        .to_string();
+
+        let item2 = json!({
+            "type": "response.output_item.done",
+            "item": {
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "World"}]
+            }
+        })
+        .to_string();
+
+        let completed = json!({
+            "type": "response.completed",
+            "response": { "id": "resp1" }
+        })
+        .to_string();
+
+        let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
+        let sse2 = format!("event: response.output_item.done\ndata: {item2}\n\n");
+        let sse3 = format!("event: response.completed\ndata: {completed}\n\n");
+
+        let provider = ModelProviderInfo {
+            name: "test".to_string(),
+            base_url: Some("https://test.com".to_string()),
+            env_key: Some("TEST_API_KEY".to_string()),
+            env_key_instructions: None,
+            wire_api: WireApi::Responses,
+            query_params: None,
+            http_headers: None,
+            env_http_headers: None,
+            request_max_retries: Some(0),
+            stream_max_retries: Some(0),
+            stream_idle_timeout_ms: Some(1000),
+            requires_auth: false,
+        };
+
+        let events = collect_events(
+            &[sse1.as_bytes(), sse2.as_bytes(), sse3.as_bytes()],
+            provider,
+        )
+        .await;
+
+        assert_eq!(events.len(), 3);
+
+        matches!(
+            &events[0],
+            Ok(ResponseEvent::OutputItemDone(ResponseItem::Message { role, .. }))
+                if role == "assistant"
+        );
+
+        matches!(
+            &events[1],
+            Ok(ResponseEvent::OutputItemDone(ResponseItem::Message { role, .. }))
+                if role == "assistant"
+        );
+
+        match &events[2] {
+            Ok(ResponseEvent::Completed {
+                response_id,
+                token_usage,
+            }) => {
+                assert_eq!(response_id, "resp1");
+                assert!(token_usage.is_none());
+            }
+            other => panic!("unexpected third event: {other:?}"),
+        }
+    }
+
+    #[tokio::test]
+    async fn error_when_missing_completed() {
+        let item1 = json!({
+            "type": "response.output_item.done",
+            "item": {
+                "type": "message",
+                "role": "assistant",
+                "content": [{"type": "output_text", "text": "Hello"}]
+            }
+        })
+        .to_string();
+
+        let sse1 = format!("event: response.output_item.done\ndata: {item1}\n\n");
+        let provider = ModelProviderInfo {
+            name: "test".to_string(),
+            base_url: Some("https://test.com".to_string()),
+            env_key: Some("TEST_API_KEY".to_string()),
+            env_key_instructions: None,
+            wire_api: WireApi::Responses,
+            query_params: None,
+            http_headers: None,
+            env_http_headers: None,
+            request_max_retries: Some(0),
+            stream_max_retries: Some(0),
+            stream_idle_timeout_ms: Some(1000),
+            requires_auth: false,
+        };
+
+        let events = collect_events(&[sse1.as_bytes()], provider).await;
+
+        assert_eq!(events.len(), 2);
+
+        matches!(events[0], Ok(ResponseEvent::OutputItemDone(_)));
+
+        match &events[1] {
+            Err(CodexErr::Stream(msg)) => {
+                assert_eq!(msg, "stream closed before response.completed")
+            }
+            other => panic!("unexpected second event: {other:?}"),
+        }
+    }
+
+    // ────────────────────────────
+    // Table-driven test from `main`
+    // ────────────────────────────
+
+    /// Verifies that the adapter produces the right `ResponseEvent` for a
+    /// variety of incoming `type` values.
     #[tokio::test]
     async fn table_driven_event_kinds() {
         struct TestCase {
@@ -441,11 +716,9 @@ mod tests {
         fn is_created(ev: &ResponseEvent) -> bool {
             matches!(ev, ResponseEvent::Created)
         }
-
         fn is_output(ev: &ResponseEvent) -> bool {
             matches!(ev, ResponseEvent::OutputItemDone(_))
         }
-
         fn is_completed(ev: &ResponseEvent) -> bool {
             matches!(ev, ResponseEvent::Completed { .. })
         }
@@ -498,9 +771,29 @@ mod tests {
         for case in cases {
             let mut evs = vec![case.event];
             evs.push(completed.clone());
-            let out = run_sse(evs).await;
+
+            let provider = ModelProviderInfo {
+                name: "test".to_string(),
+                base_url: Some("https://test.com".to_string()),
+                env_key: Some("TEST_API_KEY".to_string()),
+                env_key_instructions: None,
+                wire_api: WireApi::Responses,
+                query_params: None,
+                http_headers: None,
+                env_http_headers: None,
+                request_max_retries: Some(0),
+                stream_max_retries: Some(0),
+                stream_idle_timeout_ms: Some(1000),
+                requires_auth: false,
+            };
+
+            let out = run_sse(evs, provider).await;
             assert_eq!(out.len(), case.expected_len, "case {}", case.name);
-            assert!((case.expect_first)(&out[0]), "case {}", case.name);
+            assert!(
+                (case.expect_first)(&out[0]),
+                "first event mismatch in case {}",
+                case.name
+            );
         }
     }
 }

codex-rs/core/src/client_common.rs

@@ -22,8 +22,6 @@ const BASE_INSTRUCTIONS: &str = include_str!("../prompt.md");
 pub struct Prompt {
     /// Conversation context input items.
     pub input: Vec<ResponseItem>,
-    /// Optional previous response ID (when storage is enabled).
-    pub prev_id: Option<String>,
     /// Optional instructions from the user to amend to the built-in agent
     /// instructions.
     pub user_instructions: Option<String>,
@@ -34,14 +32,18 @@ pub struct Prompt {
     /// the "fully qualified" tool name (i.e., prefixed with the server name),
     /// which should be reported to the model in place of Tool::name.
     pub extra_tools: HashMap<String, mcp_types::Tool>,
+
+    /// Optional override for the built-in BASE_INSTRUCTIONS.
+    pub base_instructions_override: Option<String>,
 }
 
 impl Prompt {
     pub(crate) fn get_full_instructions(&self, model: &str) -> Cow<'_, str> {
-        let mut sections: Vec<&str> = vec![BASE_INSTRUCTIONS];
-        if let Some(ref user) = self.user_instructions {
-            sections.push(user);
-        }
+        let base = self
+            .base_instructions_override
+            .as_deref()
+            .unwrap_or(BASE_INSTRUCTIONS);
+        let mut sections: Vec<&str> = vec![base];
         if model.starts_with("gpt-4.1") {
             sections.push(APPLY_PATCH_TOOL_INSTRUCTIONS);
         }
@@ -57,6 +59,8 @@ pub enum ResponseEvent {
         response_id: String,
         token_usage: Option<TokenUsage>,
     },
+    OutputTextDelta(String),
+    ReasoningSummaryDelta(String),
 }
 
 #[derive(Debug, Serialize)]
@@ -124,11 +128,10 @@ pub(crate) struct ResponsesApiRequest<'a> {
     pub(crate) tool_choice: &'static str,
     pub(crate) parallel_tool_calls: bool,
     pub(crate) reasoning: Option<Reasoning>,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub(crate) previous_response_id: Option<String>,
     /// true when using the Responses API.
     pub(crate) store: bool,
     pub(crate) stream: bool,
+    pub(crate) include: Vec<String>,
 }
 
 use crate::config::Config;
@@ -182,3 +185,19 @@ impl Stream for ResponseStream {
         self.rx_event.poll_recv(cx)
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn get_full_instructions_no_user_content() {
+        let prompt = Prompt {
+            user_instructions: Some("custom instruction".to_string()),
+            ..Default::default()
+        };
+        let expected = format!("{BASE_INSTRUCTIONS}\n{APPLY_PATCH_TOOL_INSTRUCTIONS}");
+        let full = prompt.get_full_instructions("gpt-4.1");
+        assert_eq!(full, expected);
+    }
+}

codex-rs/core/src/codex_wrapper.rs

@@ -1,20 +1,37 @@
 use std::sync::Arc;
 
 use crate::Codex;
+use crate::CodexSpawnOk;
 use crate::config::Config;
 use crate::protocol::Event;
 use crate::protocol::EventMsg;
 use crate::util::notify_on_sigint;
+use codex_login::load_auth;
 use tokio::sync::Notify;
+use uuid::Uuid;
+
+/// Represents an active Codex conversation, including the first event
+/// (which is [`EventMsg::SessionConfigured`]).
+pub struct CodexConversation {
+    pub codex: Codex,
+    pub session_id: Uuid,
+    pub session_configured: Event,
+    pub ctrl_c: Arc<Notify>,
+}
 
 /// Spawn a new [`Codex`] and initialize the session.
 ///
 /// Returns the wrapped [`Codex`] **and** the `SessionInitialized` event that
 /// is received as a response to the initial `ConfigureSession` submission so
 /// that callers can surface the information to the UI.
-pub async fn init_codex(config: Config) -> anyhow::Result<(Codex, Event, Arc<Notify>)> {
+pub async fn init_codex(config: Config) -> anyhow::Result<CodexConversation> {
     let ctrl_c = notify_on_sigint();
-    let (codex, init_id) = Codex::spawn(config, ctrl_c.clone()).await?;
+    let auth = load_auth(&config.codex_home)?;
+    let CodexSpawnOk {
+        codex,
+        init_id,
+        session_id,
+    } = Codex::spawn(config, auth, ctrl_c.clone()).await?;
 
     // The first event must be `SessionInitialized`. Validate and forward it to
     // the caller so that they can display it in the conversation history.
@@ -33,5 +50,10 @@ pub async fn init_codex(config: Config) -> anyhow::Result<(Codex, Event, Arc<Not
         ));
     }
 
-    Ok((codex, event, ctrl_c))
+    Ok(CodexConversation {
+        codex,
+        session_id,
+        session_configured: event,
+        ctrl_c,
+    })
 }

codex-rs/core/src/config.rs

@@ -67,7 +67,10 @@ pub struct Config {
     pub disable_response_storage: bool,
 
     /// User-provided instructions from instructions.md.
-    pub instructions: Option<String>,
+    pub user_instructions: Option<String>,
+
+    /// Base instructions override.
+    pub base_instructions: Option<String>,
 
     /// Optional external notifier command. When set, Codex will spawn this
     /// program after each completed *turn* (i.e. when the agent finishes
@@ -141,6 +144,12 @@ pub struct Config {
 
     /// Base URL for requests to ChatGPT (as opposed to the OpenAI API).
     pub chatgpt_base_url: String,
+
+    /// Experimental rollout resume path (absolute path to .jsonl; undocumented).
+    pub experimental_resume: Option<PathBuf>,
+
+    /// Include an experimental plan tool that the model can use to update its current plan and status of each step.
+    pub include_plan_tool: bool,
 }
 
 impl Config {
@@ -329,6 +338,12 @@ pub struct ConfigToml {
 
     /// Base URL for requests to ChatGPT (as opposed to the OpenAI API).
     pub chatgpt_base_url: Option<String>,
+
+    /// Experimental rollout resume path (absolute path to .jsonl; undocumented).
+    pub experimental_resume: Option<PathBuf>,
+
+    /// Experimental path to a file whose contents replace the built-in BASE_INSTRUCTIONS.
+    pub experimental_instructions_file: Option<PathBuf>,
 }
 
 impl ConfigToml {
@@ -361,6 +376,8 @@ pub struct ConfigOverrides {
     pub model_provider: Option<String>,
     pub config_profile: Option<String>,
     pub codex_linux_sandbox_exe: Option<PathBuf>,
+    pub base_instructions: Option<String>,
+    pub include_plan_tool: Option<bool>,
 }
 
 impl Config {
@@ -371,7 +388,7 @@ impl Config {
         overrides: ConfigOverrides,
         codex_home: PathBuf,
     ) -> std::io::Result<Self> {
-        let instructions = Self::load_instructions(Some(&codex_home));
+        let user_instructions = Self::load_instructions(Some(&codex_home));
 
         // Destructure ConfigOverrides fully to ensure all overrides are applied.
         let ConfigOverrides {
@@ -382,6 +399,8 @@ impl Config {
             model_provider,
             config_profile: config_profile_key,
             codex_linux_sandbox_exe,
+            base_instructions,
+            include_plan_tool,
         } = overrides;
 
         let config_profile = match config_profile_key.as_ref().or(cfg.profile.as_ref()) {
@@ -456,6 +475,18 @@ impl Config {
                 .as_ref()
                 .map(|info| info.max_output_tokens)
         });
+
+        let experimental_resume = cfg.experimental_resume;
+
+        // Load base instructions override from a file if specified. If the
+        // path is relative, resolve it against the effective cwd so the
+        // behaviour matches other path-like config values.
+        let file_base_instructions = Self::get_base_instructions(
+            cfg.experimental_instructions_file.as_ref(),
+            &resolved_cwd,
+        )?;
+        let base_instructions = base_instructions.or(file_base_instructions);
+
         let config = Self {
             model,
             model_context_window,
@@ -474,7 +505,8 @@ impl Config {
                 .or(cfg.disable_response_storage)
                 .unwrap_or(false),
             notify: cfg.notify,
-            instructions,
+            user_instructions,
+            base_instructions,
             mcp_servers: cfg.mcp_servers,
             model_providers,
             project_doc_max_bytes: cfg.project_doc_max_bytes.unwrap_or(PROJECT_DOC_MAX_BYTES),
@@ -503,6 +535,9 @@ impl Config {
                 .chatgpt_base_url
                 .or(cfg.chatgpt_base_url)
                 .unwrap_or("https://chatgpt.com/backend-api/".to_string()),
+
+            experimental_resume,
+            include_plan_tool: include_plan_tool.unwrap_or(false),
         };
         Ok(config)
     }
@@ -523,6 +558,48 @@ impl Config {
             }
         })
     }
+
+    fn get_base_instructions(
+        path: Option<&PathBuf>,
+        cwd: &Path,
+    ) -> std::io::Result<Option<String>> {
+        let p = match path.as_ref() {
+            None => return Ok(None),
+            Some(p) => p,
+        };
+
+        // Resolve relative paths against the provided cwd to make CLI
+        // overrides consistent regardless of where the process was launched
+        // from.
+        let full_path = if p.is_relative() {
+            cwd.join(p)
+        } else {
+            p.to_path_buf()
+        };
+
+        let contents = std::fs::read_to_string(&full_path).map_err(|e| {
+            std::io::Error::new(
+                e.kind(),
+                format!(
+                    "failed to read experimental instructions file {}: {e}",
+                    full_path.display()
+                ),
+            )
+        })?;
+
+        let s = contents.trim().to_string();
+        if s.is_empty() {
+            Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                format!(
+                    "experimental instructions file is empty: {}",
+                    full_path.display()
+                ),
+            ))
+        } else {
+            Ok(Some(s))
+        }
+    }
 }
 
 fn default_model() -> String {
@@ -537,7 +614,7 @@ fn default_model() -> String {
 ///   function will Err if the path does not exist.
 /// - If `CODEX_HOME` is not set, this function does not verify that the
 ///   directory exists.
-fn find_codex_home() -> std::io::Result<PathBuf> {
+pub fn find_codex_home() -> std::io::Result<PathBuf> {
     // Honor the `CODEX_HOME` environment variable when it is set to allow users
     // (and tests) to override the default location.
     if let Ok(val) = std::env::var("CODEX_HOME") {
@@ -691,6 +768,9 @@ name = "OpenAI using Chat Completions"
 base_url = "https://api.openai.com/v1"
 env_key = "OPENAI_API_KEY"
 wire_api = "chat"
+request_max_retries = 4            # retry failed HTTP requests
+stream_max_retries = 10            # retry dropped SSE streams
+stream_idle_timeout_ms = 300000    # 5m idle timeout
 
 [profiles.o3]
 model = "o3"
@@ -724,13 +804,17 @@ disable_response_storage = true
 
         let openai_chat_completions_provider = ModelProviderInfo {
             name: "OpenAI using Chat Completions".to_string(),
-            base_url: "https://api.openai.com/v1".to_string(),
+            base_url: Some("https://api.openai.com/v1".to_string()),
             env_key: Some("OPENAI_API_KEY".to_string()),
             wire_api: crate::WireApi::Chat,
             env_key_instructions: None,
             query_params: None,
             http_headers: None,
             env_http_headers: None,
+            request_max_retries: Some(4),
+            stream_max_retries: Some(10),
+            stream_idle_timeout_ms: Some(300_000),
+            requires_auth: false,
         };
         let model_provider_map = {
             let mut model_provider_map = built_in_model_providers();
@@ -761,7 +845,7 @@ disable_response_storage = true
     ///
     /// 1. custom command-line argument, e.g. `--model o3`
     /// 2. as part of a profile, where the `--profile` is specified via a CLI
-    ///    (or in the config file itelf)
+    ///    (or in the config file itself)
     /// 3. as an entry in `config.toml`, e.g. `model = "o3"`
     /// 4. the default value for a required field defined in code, e.g.,
     ///    `crate::flags::OPENAI_DEFAULT_MODEL`
@@ -793,7 +877,7 @@ disable_response_storage = true
                 sandbox_policy: SandboxPolicy::new_read_only_policy(),
                 shell_environment_policy: ShellEnvironmentPolicy::default(),
                 disable_response_storage: false,
-                instructions: None,
+                user_instructions: None,
                 notify: None,
                 cwd: fixture.cwd(),
                 mcp_servers: HashMap::new(),
@@ -810,6 +894,9 @@ disable_response_storage = true
                 model_reasoning_summary: ReasoningSummary::Detailed,
                 model_supports_reasoning_summaries: false,
                 chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
+                experimental_resume: None,
+                base_instructions: None,
+                include_plan_tool: false,
             },
             o3_profile_config
         );
@@ -840,7 +927,7 @@ disable_response_storage = true
             sandbox_policy: SandboxPolicy::new_read_only_policy(),
             shell_environment_policy: ShellEnvironmentPolicy::default(),
             disable_response_storage: false,
-            instructions: None,
+            user_instructions: None,
             notify: None,
             cwd: fixture.cwd(),
             mcp_servers: HashMap::new(),
@@ -857,6 +944,9 @@ disable_response_storage = true
             model_reasoning_summary: ReasoningSummary::default(),
             model_supports_reasoning_summaries: false,
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
+            experimental_resume: None,
+            base_instructions: None,
+            include_plan_tool: false,
         };
 
         assert_eq!(expected_gpt3_profile_config, gpt3_profile_config);
@@ -902,7 +992,7 @@ disable_response_storage = true
             sandbox_policy: SandboxPolicy::new_read_only_policy(),
             shell_environment_policy: ShellEnvironmentPolicy::default(),
             disable_response_storage: true,
-            instructions: None,
+            user_instructions: None,
             notify: None,
             cwd: fixture.cwd(),
             mcp_servers: HashMap::new(),
@@ -919,6 +1009,9 @@ disable_response_storage = true
             model_reasoning_summary: ReasoningSummary::default(),
             model_supports_reasoning_summaries: false,
             chatgpt_base_url: "https://chatgpt.com/backend-api/".to_string(),
+            experimental_resume: None,
+            base_instructions: None,
+            include_plan_tool: false,
         };
 
         assert_eq!(expected_zdr_profile_config, zdr_profile_config);

codex-rs/core/src/config_types.rs

@@ -76,22 +76,9 @@ pub enum HistoryPersistence {
 
 /// Collection of settings that are specific to the TUI.
 #[derive(Deserialize, Debug, Clone, PartialEq, Default)]
-pub struct Tui {
-    /// By default, mouse capture is enabled in the TUI so that it is possible
-    /// to scroll the conversation history with a mouse. This comes at the cost
-    /// of not being able to use the mouse to select text in the TUI.
-    /// (Most terminals support a modifier key to allow this. For example,
-    /// text selection works in iTerm if you hold down the `Option` key while
-    /// clicking and dragging.)
-    ///
-    /// Setting this option to `true` disables mouse capture, so scrolling with
-    /// the mouse is not possible, though the keyboard shortcuts e.g. `b` and
-    /// `space` still work. This allows the user to select text in the TUI
-    /// using the mouse without needing to hold down a modifier key.
-    pub disable_mouse_capture: bool,
-}
+pub struct Tui {}
 
-#[derive(Deserialize, Debug, Clone, Copy, PartialEq, Default)]
+#[derive(Deserialize, Debug, Clone, Copy, PartialEq, Default, Serialize)]
 #[serde(rename_all = "kebab-case")]
 pub enum SandboxMode {
     #[serde(rename = "read-only")]
@@ -143,6 +130,8 @@ pub struct ShellEnvironmentPolicyToml {
 
     /// List of regular expressions.
     pub include_only: Option<Vec<String>>,
+
+    pub experimental_use_profile: Option<bool>,
 }
 
 pub type EnvironmentVariablePattern = WildMatchPattern<'*', '?'>;
@@ -171,6 +160,9 @@ pub struct ShellEnvironmentPolicy {
 
     /// Environment variable names to retain in the environment.
     pub include_only: Vec<EnvironmentVariablePattern>,
+
+    /// If true, the shell profile will be used to run the command.
+    pub use_profile: bool,
 }
 
 impl From<ShellEnvironmentPolicyToml> for ShellEnvironmentPolicy {
@@ -190,6 +182,7 @@ impl From<ShellEnvironmentPolicyToml> for ShellEnvironmentPolicy {
             .into_iter()
             .map(|s| EnvironmentVariablePattern::new_case_insensitive(&s))
             .collect();
+        let use_profile = toml.experimental_use_profile.unwrap_or(false);
 
         Self {
             inherit,
@@ -197,6 +190,7 @@ impl From<ShellEnvironmentPolicyToml> for ShellEnvironmentPolicy {
             exclude,
             r#set,
             include_only,
+            use_profile,
         }
     }
 }

codex-rs/core/src/conversation_history.rs

@@ -1,12 +1,7 @@
 use crate::models::ResponseItem;
 
-/// Transcript of conversation history that is needed:
-/// - for ZDR clients for which previous_response_id is not available, so we
-///   must include the transcript with every API call. This must include each
-///   `function_call` and its corresponding `function_call_output`.
-/// - for clients using the "chat completions" API as opposed to the
-///   "responses" API.
-#[derive(Debug, Clone)]
+/// Transcript of conversation history
+#[derive(Debug, Clone, Default)]
 pub(crate) struct ConversationHistory {
     /// The oldest items are at the beginning of the vector.
     items: Vec<ResponseItem>,
@@ -44,7 +39,8 @@ fn is_api_message(message: &ResponseItem) -> bool {
         ResponseItem::Message { role, .. } => role.as_str() != "system",
         ResponseItem::FunctionCallOutput { .. }
         | ResponseItem::FunctionCall { .. }
-        | ResponseItem::LocalShellCall { .. } => true,
-        ResponseItem::Reasoning { .. } | ResponseItem::Other => false,
+        | ResponseItem::LocalShellCall { .. }
+        | ResponseItem::Reasoning { .. } => true,
+        ResponseItem::Other => false,
     }
 }

codex-rs/core/src/exec.rs

@@ -17,6 +17,7 @@ use tokio::io::BufReader;
 use tokio::process::Child;
 use tokio::process::Command;
 use tokio::sync::Notify;
+use tracing::trace;
 
 use crate::error::CodexErr;
 use crate::error::Result;
@@ -82,7 +83,8 @@ pub async fn process_exec_tool_call(
 ) -> Result<ExecToolCallOutput> {
     let start = Instant::now();
 
-    let raw_output_result = match sandbox_type {
+    let raw_output_result: std::result::Result<RawExecToolCallOutput, CodexErr> = match sandbox_type
+    {
         SandboxType::None => exec(params, sandbox_policy, ctrl_c).await,
         SandboxType::MacosSeatbelt => {
             let ExecParams {
@@ -372,6 +374,10 @@ async fn spawn_child_async(
     stdio_policy: StdioPolicy,
     env: HashMap<String, String>,
 ) -> std::io::Result<Child> {
+    trace!(
+        "spawn_child_async: {program:?} {args:?} {arg0:?} {cwd:?} {sandbox_policy:?} {stdio_policy:?} {env:?}"
+    );
+
     let mut cmd = Command::new(&program);
     #[cfg(unix)]
     cmd.arg0(arg0.map_or_else(|| program.to_string_lossy().to_string(), String::from));
@@ -384,6 +390,31 @@ async fn spawn_child_async(
         cmd.env(CODEX_SANDBOX_NETWORK_DISABLED_ENV_VAR, "1");
     }
 
+    // If this Codex process dies (including being killed via SIGKILL), we want
+    // any child processes that were spawned as part of a `"shell"` tool call
+    // to also be terminated.
+
+    // This relies on prctl(2), so it only works on Linux.
+    #[cfg(target_os = "linux")]
+    unsafe {
+        cmd.pre_exec(|| {
+            // This prctl call effectively requests, "deliver SIGTERM when my
+            // current parent dies."
+            if libc::prctl(libc::PR_SET_PDEATHSIG, libc::SIGTERM) == -1 {
+                return Err(io::Error::last_os_error());
+            }
+
+            // Though if there was a race condition and this pre_exec() block is
+            // run _after_ the parent (i.e., the Codex process) has already
+            // exited, then the parent is the _init_ process (which will never
+            // die), so we should just terminate the child process now.
+            if libc::getppid() == 1 {
+                libc::raise(libc::SIGTERM);
+            }
+            Ok(())
+        });
+    }
+
     match stdio_policy {
         StdioPolicy::RedirectForShellTool => {
             // Do not create a file descriptor for stdin because otherwise some

codex-rs/core/src/flags.rs

@@ -11,14 +11,6 @@ env_flags! {
     pub OPENAI_TIMEOUT_MS: Duration = Duration::from_millis(300_000), |value| {
         value.parse().map(Duration::from_millis)
     };
-    pub OPENAI_REQUEST_MAX_RETRIES: u64 = 4;
-    pub OPENAI_STREAM_MAX_RETRIES: u64 = 10;
-
-    // We generally don't want to disconnect; this updates the timeout to be five minutes
-    // which matches the upstream typescript codex impl.
-    pub OPENAI_STREAM_IDLE_TIMEOUT_MS: Duration = Duration::from_millis(300_000), |value| {
-        value.parse().map(Duration::from_millis)
-    };
 
     /// Fixture path for offline tests (see client.rs).
     pub CODEX_RS_SSE_FIXTURE: Option<&str> = None;

codex-rs/core/src/git_info.rs

@@ -0,0 +1,307 @@
+use std::path::Path;
+
+use serde::Deserialize;
+use serde::Serialize;
+use tokio::process::Command;
+use tokio::time::Duration as TokioDuration;
+use tokio::time::timeout;
+
+/// Timeout for git commands to prevent freezing on large repositories
+const GIT_COMMAND_TIMEOUT: TokioDuration = TokioDuration::from_secs(5);
+
+#[derive(Serialize, Deserialize, Clone)]
+pub struct GitInfo {
+    /// Current commit hash (SHA)
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub commit_hash: Option<String>,
+    /// Current branch name
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub branch: Option<String>,
+    /// Repository URL (if available from remote)
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub repository_url: Option<String>,
+}
+
+/// Collect git repository information from the given working directory using command-line git.
+/// Returns None if no git repository is found or if git operations fail.
+/// Uses timeouts to prevent freezing on large repositories.
+/// All git commands (except the initial repo check) run in parallel for better performance.
+pub async fn collect_git_info(cwd: &Path) -> Option<GitInfo> {
+    // Check if we're in a git repository first
+    let is_git_repo = run_git_command_with_timeout(&["rev-parse", "--git-dir"], cwd)
+        .await?
+        .status
+        .success();
+
+    if !is_git_repo {
+        return None;
+    }
+
+    // Run all git info collection commands in parallel
+    let (commit_result, branch_result, url_result) = tokio::join!(
+        run_git_command_with_timeout(&["rev-parse", "HEAD"], cwd),
+        run_git_command_with_timeout(&["rev-parse", "--abbrev-ref", "HEAD"], cwd),
+        run_git_command_with_timeout(&["remote", "get-url", "origin"], cwd)
+    );
+
+    let mut git_info = GitInfo {
+        commit_hash: None,
+        branch: None,
+        repository_url: None,
+    };
+
+    // Process commit hash
+    if let Some(output) = commit_result {
+        if output.status.success() {
+            if let Ok(hash) = String::from_utf8(output.stdout) {
+                git_info.commit_hash = Some(hash.trim().to_string());
+            }
+        }
+    }
+
+    // Process branch name
+    if let Some(output) = branch_result {
+        if output.status.success() {
+            if let Ok(branch) = String::from_utf8(output.stdout) {
+                let branch = branch.trim();
+                if branch != "HEAD" {
+                    git_info.branch = Some(branch.to_string());
+                }
+            }
+        }
+    }
+
+    // Process repository URL
+    if let Some(output) = url_result {
+        if output.status.success() {
+            if let Ok(url) = String::from_utf8(output.stdout) {
+                git_info.repository_url = Some(url.trim().to_string());
+            }
+        }
+    }
+
+    Some(git_info)
+}
+
+/// Run a git command with a timeout to prevent blocking on large repositories
+async fn run_git_command_with_timeout(args: &[&str], cwd: &Path) -> Option<std::process::Output> {
+    let result = timeout(
+        GIT_COMMAND_TIMEOUT,
+        Command::new("git").args(args).current_dir(cwd).output(),
+    )
+    .await;
+
+    match result {
+        Ok(Ok(output)) => Some(output),
+        _ => None, // Timeout or error
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    #![allow(clippy::expect_used)]
+    #![allow(clippy::unwrap_used)]
+
+    use super::*;
+
+    use std::fs;
+    use std::path::PathBuf;
+    use tempfile::TempDir;
+
+    // Helper function to create a test git repository
+    async fn create_test_git_repo(temp_dir: &TempDir) -> PathBuf {
+        let repo_path = temp_dir.path().to_path_buf();
+
+        // Initialize git repo
+        Command::new("git")
+            .args(["init"])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to init git repo");
+
+        // Configure git user (required for commits)
+        Command::new("git")
+            .args(["config", "user.name", "Test User"])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to set git user name");
+
+        Command::new("git")
+            .args(["config", "user.email", "test@example.com"])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to set git user email");
+
+        // Create a test file and commit it
+        let test_file = repo_path.join("test.txt");
+        fs::write(&test_file, "test content").expect("Failed to write test file");
+
+        Command::new("git")
+            .args(["add", "."])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to add files");
+
+        Command::new("git")
+            .args(["commit", "-m", "Initial commit"])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to commit");
+
+        repo_path
+    }
+
+    #[tokio::test]
+    async fn test_collect_git_info_non_git_directory() {
+        let temp_dir = TempDir::new().expect("Failed to create temp dir");
+        let result = collect_git_info(temp_dir.path()).await;
+        assert!(result.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_collect_git_info_git_repository() {
+        let temp_dir = TempDir::new().expect("Failed to create temp dir");
+        let repo_path = create_test_git_repo(&temp_dir).await;
+
+        let git_info = collect_git_info(&repo_path)
+            .await
+            .expect("Should collect git info from repo");
+
+        // Should have commit hash
+        assert!(git_info.commit_hash.is_some());
+        let commit_hash = git_info.commit_hash.unwrap();
+        assert_eq!(commit_hash.len(), 40); // SHA-1 hash should be 40 characters
+        assert!(commit_hash.chars().all(|c| c.is_ascii_hexdigit()));
+
+        // Should have branch (likely "main" or "master")
+        assert!(git_info.branch.is_some());
+        let branch = git_info.branch.unwrap();
+        assert!(branch == "main" || branch == "master");
+
+        // Repository URL might be None for local repos without remote
+        // This is acceptable behavior
+    }
+
+    #[tokio::test]
+    async fn test_collect_git_info_with_remote() {
+        let temp_dir = TempDir::new().expect("Failed to create temp dir");
+        let repo_path = create_test_git_repo(&temp_dir).await;
+
+        // Add a remote origin
+        Command::new("git")
+            .args([
+                "remote",
+                "add",
+                "origin",
+                "https://github.com/example/repo.git",
+            ])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to add remote");
+
+        let git_info = collect_git_info(&repo_path)
+            .await
+            .expect("Should collect git info from repo");
+
+        // Should have repository URL
+        assert_eq!(
+            git_info.repository_url,
+            Some("https://github.com/example/repo.git".to_string())
+        );
+    }
+
+    #[tokio::test]
+    async fn test_collect_git_info_detached_head() {
+        let temp_dir = TempDir::new().expect("Failed to create temp dir");
+        let repo_path = create_test_git_repo(&temp_dir).await;
+
+        // Get the current commit hash
+        let output = Command::new("git")
+            .args(["rev-parse", "HEAD"])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to get HEAD");
+        let commit_hash = String::from_utf8(output.stdout).unwrap().trim().to_string();
+
+        // Checkout the commit directly (detached HEAD)
+        Command::new("git")
+            .args(["checkout", &commit_hash])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to checkout commit");
+
+        let git_info = collect_git_info(&repo_path)
+            .await
+            .expect("Should collect git info from repo");
+
+        // Should have commit hash
+        assert!(git_info.commit_hash.is_some());
+        // Branch should be None for detached HEAD (since rev-parse --abbrev-ref HEAD returns "HEAD")
+        assert!(git_info.branch.is_none());
+    }
+
+    #[tokio::test]
+    async fn test_collect_git_info_with_branch() {
+        let temp_dir = TempDir::new().expect("Failed to create temp dir");
+        let repo_path = create_test_git_repo(&temp_dir).await;
+
+        // Create and checkout a new branch
+        Command::new("git")
+            .args(["checkout", "-b", "feature-branch"])
+            .current_dir(&repo_path)
+            .output()
+            .await
+            .expect("Failed to create branch");
+
+        let git_info = collect_git_info(&repo_path)
+            .await
+            .expect("Should collect git info from repo");
+
+        // Should have the new branch name
+        assert_eq!(git_info.branch, Some("feature-branch".to_string()));
+    }
+
+    #[test]
+    fn test_git_info_serialization() {
+        let git_info = GitInfo {
+            commit_hash: Some("abc123def456".to_string()),
+            branch: Some("main".to_string()),
+            repository_url: Some("https://github.com/example/repo.git".to_string()),
+        };
+
+        let json = serde_json::to_string(&git_info).expect("Should serialize GitInfo");
+        let parsed: serde_json::Value = serde_json::from_str(&json).expect("Should parse JSON");
+
+        assert_eq!(parsed["commit_hash"], "abc123def456");
+        assert_eq!(parsed["branch"], "main");
+        assert_eq!(
+            parsed["repository_url"],
+            "https://github.com/example/repo.git"
+        );
+    }
+
+    #[test]
+    fn test_git_info_serialization_with_nones() {
+        let git_info = GitInfo {
+            commit_hash: None,
+            branch: None,
+            repository_url: None,
+        };
+
+        let json = serde_json::to_string(&git_info).expect("Should serialize GitInfo");
+        let parsed: serde_json::Value = serde_json::from_str(&json).expect("Should parse JSON");
+
+        // Fields with None values should be omitted due to skip_serializing_if
+        assert!(!parsed.as_object().unwrap().contains_key("commit_hash"));
+        assert!(!parsed.as_object().unwrap().contains_key("branch"));
+        assert!(!parsed.as_object().unwrap().contains_key("repository_url"));
+    }
+}

codex-rs/core/src/is_safe_command.rs

@@ -1,31 +1,57 @@
-use tree_sitter::Parser;
-use tree_sitter::Tree;
-use tree_sitter_bash::LANGUAGE as BASH;
+use crate::bash::try_parse_bash;
+use crate::bash::try_parse_word_only_commands_sequence;
 
 pub fn is_known_safe_command(command: &[String]) -> bool {
     if is_safe_to_call_with_exec(command) {
         return true;
     }
 
-    // TODO(mbolin): Also support safe commands that are piped together such
-    // as `cat foo | wc -l`.
-    matches!(
-        command,
-        [bash, flag, script]
-            if bash == "bash"
-            && flag == "-lc"
-            && try_parse_bash(script).and_then(|tree|
-                try_parse_single_word_only_command(&tree, script)).is_some_and(|parsed_bash_command| is_safe_to_call_with_exec(&parsed_bash_command))
-    )
+    // Support `bash -lc "..."` where the script consists solely of one or
+    // more "plain" commands (only bare words / quoted strings) combined with
+    // a conservative allow‑list of shell operators that themselves do not
+    // introduce side effects ( "&&", "||", ";", and "|" ). If every
+    // individual command in the script is itself a known‑safe command, then
+    // the composite expression is considered safe.
+    if let [bash, flag, script] = command {
+        if bash == "bash" && flag == "-lc" {
+            if let Some(tree) = try_parse_bash(script) {
+                if let Some(all_commands) = try_parse_word_only_commands_sequence(&tree, script) {
+                    if !all_commands.is_empty()
+                        && all_commands
+                            .iter()
+                            .all(|cmd| is_safe_to_call_with_exec(cmd))
+                    {
+                        return true;
+                    }
+                }
+            }
+        }
+    }
+
+    false
 }
 
 fn is_safe_to_call_with_exec(command: &[String]) -> bool {
     let cmd0 = command.first().map(String::as_str);
 
     match cmd0 {
+        #[rustfmt::skip]
         Some(
-            "cat" | "cd" | "echo" | "grep" | "head" | "ls" | "pwd" | "rg" | "tail" | "wc" | "which",
-        ) => true,
+            "cat" |
+            "cd" |
+            "echo" |
+            "false" |
+            "grep" |
+            "head" |
+            "ls" |
+            "nl" |
+            "pwd" |
+            "tail" |
+            "true" |
+            "wc" |
+            "which") => {
+            true
+        },
 
         Some("find") => {
             // Certain options to `find` can delete files, write to files, or
@@ -46,6 +72,29 @@ fn is_safe_to_call_with_exec(command: &[String]) -> bool {
                 .any(|arg| UNSAFE_FIND_OPTIONS.contains(&arg.as_str()))
         }
 
+        // Ripgrep
+        Some("rg") => {
+            const UNSAFE_RIPGREP_OPTIONS_WITH_ARGS: &[&str] = &[
+                // Takes an arbitrary command that is executed for each match.
+                "--pre",
+                // Takes a command that can be used to obtain the local hostname.
+                "--hostname-bin",
+            ];
+            const UNSAFE_RIPGREP_OPTIONS_WITHOUT_ARGS: &[&str] = &[
+                // Calls out to other decompression tools, so do not auto-approve
+                // out of an abundance of caution.
+                "--search-zip",
+                "-z",
+            ];
+
+            !command.iter().any(|arg| {
+                UNSAFE_RIPGREP_OPTIONS_WITHOUT_ARGS.contains(&arg.as_str())
+                    || UNSAFE_RIPGREP_OPTIONS_WITH_ARGS
+                        .iter()
+                        .any(|&opt| arg == opt || arg.starts_with(&format!("{opt}=")))
+            })
+        }
+
         // Git
         Some("git") => matches!(
             command.get(1).map(String::as_str),
@@ -72,90 +121,7 @@ fn is_safe_to_call_with_exec(command: &[String]) -> bool {
     }
 }
 
-fn try_parse_bash(bash_lc_arg: &str) -> Option<Tree> {
-    let lang = BASH.into();
-    let mut parser = Parser::new();
-    #[expect(clippy::expect_used)]
-    parser.set_language(&lang).expect("load bash grammar");
-
-    let old_tree: Option<&Tree> = None;
-    parser.parse(bash_lc_arg, old_tree)
-}
-
-/// If `tree` represents a single Bash command whose name and every argument is
-/// an ordinary `word`, return those words in order; otherwise, return `None`.
-///
-/// `src` must be the exact source string that was parsed into `tree`, so we can
-/// extract the text for every node.
-pub fn try_parse_single_word_only_command(tree: &Tree, src: &str) -> Option<Vec<String>> {
-    // Any parse error is an immediate rejection.
-    if tree.root_node().has_error() {
-        return None;
-    }
-
-    // (program …) with exactly one statement
-    let root = tree.root_node();
-    if root.kind() != "program" || root.named_child_count() != 1 {
-        return None;
-    }
-
-    let cmd = root.named_child(0)?; // (command …)
-    if cmd.kind() != "command" {
-        return None;
-    }
-
-    let mut words = Vec::new();
-    let mut cursor = cmd.walk();
-
-    for child in cmd.named_children(&mut cursor) {
-        match child.kind() {
-            // The command name node wraps one `word` child.
-            "command_name" => {
-                let word_node = child.named_child(0)?; // make sure it's only a word
-                if word_node.kind() != "word" {
-                    return None;
-                }
-                words.push(word_node.utf8_text(src.as_bytes()).ok()?.to_owned());
-            }
-            // Positional‑argument word (allowed).
-            "word" | "number" => {
-                words.push(child.utf8_text(src.as_bytes()).ok()?.to_owned());
-            }
-            "string" => {
-                if child.child_count() == 3
-                    && child.child(0)?.kind() == "\""
-                    && child.child(1)?.kind() == "string_content"
-                    && child.child(2)?.kind() == "\""
-                {
-                    words.push(child.child(1)?.utf8_text(src.as_bytes()).ok()?.to_owned());
-                } else {
-                    // Anything else means the command is *not* plain words.
-                    return None;
-                }
-            }
-            "concatenation" => {
-                // TODO: Consider things like `'ab\'a'`.
-                return None;
-            }
-            "raw_string" => {
-                // Raw string is a single word, but we need to strip the quotes.
-                let raw_string = child.utf8_text(src.as_bytes()).ok()?;
-                let stripped = raw_string
-                    .strip_prefix('\'')
-                    .and_then(|s| s.strip_suffix('\''));
-                if let Some(stripped) = stripped {
-                    words.push(stripped.to_owned());
-                } else {
-                    return None;
-                }
-            }
-            // Anything else means the command is *not* plain words.
-            _ => return None,
-        }
-    }
-
-    Some(words)
-}
+// (bash parsing helpers implemented in crate::bash)
 
 /* ----------------------------------------------------------
 Example
@@ -193,6 +159,7 @@ fn is_valid_sed_n_arg(arg: Option<&str>) -> bool {
         _ => false,
     }
 }
+
 #[cfg(test)]
 mod tests {
     #![allow(clippy::unwrap_used)]
@@ -209,6 +176,11 @@ mod tests {
         assert!(is_safe_to_call_with_exec(&vec_str(&[
             "sed", "-n", "1,5p", "file.txt"
         ])));
+        assert!(is_safe_to_call_with_exec(&vec_str(&[
+            "nl",
+            "-nrz",
+            "Cargo.toml"
+        ])));
 
         // Safe `find` command (no unsafe options).
         assert!(is_safe_to_call_with_exec(&vec_str(&[
@@ -245,6 +217,40 @@ mod tests {
         }
     }
 
+    #[test]
+    fn ripgrep_rules() {
+        // Safe ripgrep invocations – none of the unsafe flags are present.
+        assert!(is_safe_to_call_with_exec(&vec_str(&[
+            "rg",
+            "Cargo.toml",
+            "-n"
+        ])));
+
+        // Unsafe flags that do not take an argument (present verbatim).
+        for args in [
+            vec_str(&["rg", "--search-zip", "files"]),
+            vec_str(&["rg", "-z", "files"]),
+        ] {
+            assert!(
+                !is_safe_to_call_with_exec(&args),
+                "expected {args:?} to be considered unsafe due to zip-search flag",
+            );
+        }
+
+        // Unsafe flags that expect a value, provided in both split and = forms.
+        for args in [
+            vec_str(&["rg", "--pre", "pwned", "files"]),
+            vec_str(&["rg", "--pre=pwned", "files"]),
+            vec_str(&["rg", "--hostname-bin", "pwned", "files"]),
+            vec_str(&["rg", "--hostname-bin=pwned", "files"]),
+        ] {
+            assert!(
+                !is_safe_to_call_with_exec(&args),
+                "expected {args:?} to be considered unsafe due to external-command flag",
+            );
+        }
+    }
+
     #[test]
     fn bash_lc_safe_examples() {
         assert!(is_known_safe_command(&vec_str(&["bash", "-lc", "ls"])));
@@ -277,6 +283,30 @@ mod tests {
         ])));
     }
 
+    #[test]
+    fn bash_lc_safe_examples_with_operators() {
+        assert!(is_known_safe_command(&vec_str(&[
+            "bash",
+            "-lc",
+            "grep -R \"Cargo.toml\" -n || true"
+        ])));
+        assert!(is_known_safe_command(&vec_str(&[
+            "bash",
+            "-lc",
+            "ls && pwd"
+        ])));
+        assert!(is_known_safe_command(&vec_str(&[
+            "bash",
+            "-lc",
+            "echo 'hi' ; ls"
+        ])));
+        assert!(is_known_safe_command(&vec_str(&[
+            "bash",
+            "-lc",
+            "ls | wc -l"
+        ])));
+    }
+
     #[test]
     fn bash_lc_unsafe_examples() {
         assert!(
@@ -290,44 +320,29 @@ mod tests {
 
         assert!(
             !is_known_safe_command(&vec_str(&["bash", "-lc", "find . -name file.txt -delete"])),
-            "Unsafe find option should not be auto‑approved."
-        );
-    }
-
-    #[test]
-    fn test_try_parse_single_word_only_command() {
-        let script_with_single_quoted_string = "sed -n '1,5p' file.txt";
-        let parsed_words = try_parse_bash(script_with_single_quoted_string)
-            .and_then(|tree| {
-                try_parse_single_word_only_command(&tree, script_with_single_quoted_string)
-            })
-            .unwrap();
-        assert_eq!(
-            vec![
-                "sed".to_string(),
-                "-n".to_string(),
-                // Ensure the single quotes are properly removed.
-                "1,5p".to_string(),
-                "file.txt".to_string()
-            ],
-            parsed_words,
+            "Unsafe find option should not be auto-approved."
         );
 
-        let script_with_number_arg = "ls -1";
-        let parsed_words = try_parse_bash(script_with_number_arg)
-            .and_then(|tree| try_parse_single_word_only_command(&tree, script_with_number_arg))
-            .unwrap();
-        assert_eq!(vec!["ls", "-1"], parsed_words,);
+        // Disallowed because of unsafe command in sequence.
+        assert!(
+            !is_known_safe_command(&vec_str(&["bash", "-lc", "ls && rm -rf /"])),
+            "Sequence containing unsafe command must be rejected"
+        );
 
-        let script_with_double_quoted_string_with_no_funny_stuff_arg = "grep -R \"Cargo.toml\" -n";
-        let parsed_words = try_parse_bash(script_with_double_quoted_string_with_no_funny_stuff_arg)
-            .and_then(|tree| {
-                try_parse_single_word_only_command(
-                    &tree,
-                    script_with_double_quoted_string_with_no_funny_stuff_arg,
-                )
-            })
-            .unwrap();
-        assert_eq!(vec!["grep", "-R", "Cargo.toml", "-n"], parsed_words);
+        // Disallowed because of parentheses / subshell.
+        assert!(
+            !is_known_safe_command(&vec_str(&["bash", "-lc", "(ls)"])),
+            "Parentheses (subshell) are not provably safe with the current parser"
+        );
+        assert!(
+            !is_known_safe_command(&vec_str(&["bash", "-lc", "ls || (pwd && echo hi)"])),
+            "Nested parentheses are not provably safe with the current parser"
+        );
+
+        // Disallowed redirection.
+        assert!(
+            !is_known_safe_command(&vec_str(&["bash", "-lc", "ls > out.txt"])),
+            "> redirection should be rejected"
+        );
     }
 }

codex-rs/core/src/lib.rs

@@ -5,11 +5,14 @@
 // the TUI or the tracing stack).
 #![deny(clippy::print_stdout, clippy::print_stderr)]
 
+mod apply_patch;
+mod bash;
 mod chat_completions;
 mod client;
 mod client_common;
 pub mod codex;
 pub use codex::Codex;
+pub use codex::CodexSpawnOk;
 pub mod codex_wrapper;
 pub mod config;
 pub mod config_profile;
@@ -19,6 +22,7 @@ pub mod error;
 pub mod exec;
 pub mod exec_env;
 mod flags;
+pub mod git_info;
 mod is_safe_command;
 mod mcp_connection_manager;
 mod mcp_tool_call;
@@ -26,15 +30,18 @@ mod message_history;
 mod model_provider_info;
 pub use model_provider_info::ModelProviderInfo;
 pub use model_provider_info::WireApi;
+pub use model_provider_info::built_in_model_providers;
 mod models;
-pub mod openai_api_key;
 mod openai_model_info;
 mod openai_tools;
+pub mod plan_tool;
 mod project_doc;
 pub mod protocol;
 mod rollout;
 mod safety;
+pub mod shell;
 mod user_notification;
 pub mod util;
 
+pub use apply_patch::CODEX_APPLY_PATCH_ARG1;
 pub use client_common::model_supports_reasoning_summaries;

codex-rs/core/src/mcp_connection_manager.rs

@@ -7,6 +7,8 @@
 //! `"<server><MCP_TOOL_NAME_DELIMITER><tool>"` as the key.
 
 use std::collections::HashMap;
+use std::collections::HashSet;
+use std::ffi::OsString;
 use std::time::Duration;
 
 use anyhow::Context;
@@ -16,8 +18,13 @@ use codex_mcp_client::McpClient;
 use mcp_types::ClientCapabilities;
 use mcp_types::Implementation;
 use mcp_types::Tool;
+
+use serde_json::json;
+use sha1::Digest;
+use sha1::Sha1;
 use tokio::task::JoinSet;
 use tracing::info;
+use tracing::warn;
 
 use crate::config_types::McpServerConfig;
 
@@ -26,7 +33,8 @@ use crate::config_types::McpServerConfig;
 ///
 /// OpenAI requires tool names to conform to `^[a-zA-Z0-9_-]+$`, so we must
 /// choose a delimiter from this character set.
-const MCP_TOOL_NAME_DELIMITER: &str = "__OAI_CODEX_MCP__";
+const MCP_TOOL_NAME_DELIMITER: &str = "__";
+const MAX_TOOL_NAME_LENGTH: usize = 64;
 
 /// Timeout for the `tools/list` request.
 const LIST_TOOLS_TIMEOUT: Duration = Duration::from_secs(10);
@@ -35,16 +43,42 @@ const LIST_TOOLS_TIMEOUT: Duration = Duration::from_secs(10);
 /// spawned successfully.
 pub type ClientStartErrors = HashMap<String, anyhow::Error>;
 
-fn fully_qualified_tool_name(server: &str, tool: &str) -> String {
-    format!("{server}{MCP_TOOL_NAME_DELIMITER}{tool}")
+fn qualify_tools(tools: Vec<ToolInfo>) -> HashMap<String, ToolInfo> {
+    let mut used_names = HashSet::new();
+    let mut qualified_tools = HashMap::new();
+    for tool in tools {
+        let mut qualified_name = format!(
+            "{}{}{}",
+            tool.server_name, MCP_TOOL_NAME_DELIMITER, tool.tool_name
+        );
+        if qualified_name.len() > MAX_TOOL_NAME_LENGTH {
+            let mut hasher = Sha1::new();
+            hasher.update(qualified_name.as_bytes());
+            let sha1 = hasher.finalize();
+            let sha1_str = format!("{sha1:x}");
+
+            // Truncate to make room for the hash suffix
+            let prefix_len = MAX_TOOL_NAME_LENGTH - sha1_str.len();
+
+            qualified_name = format!("{}{}", &qualified_name[..prefix_len], sha1_str);
+        }
+
+        if used_names.contains(&qualified_name) {
+            warn!("skipping duplicated tool {}", qualified_name);
+            continue;
+        }
+
+        used_names.insert(qualified_name.clone());
+        qualified_tools.insert(qualified_name, tool);
+    }
+
+    qualified_tools
 }
 
-pub(crate) fn try_parse_fully_qualified_tool_name(fq_name: &str) -> Option<(String, String)> {
-    let (server, tool) = fq_name.split_once(MCP_TOOL_NAME_DELIMITER)?;
-    if server.is_empty() || tool.is_empty() {
-        return None;
-    }
-    Some((server.to_string(), tool.to_string()))
+struct ToolInfo {
+    server_name: String,
+    tool_name: String,
+    tool: Tool,
 }
 
 /// A thin wrapper around a set of running [`McpClient`] instances.
@@ -57,7 +91,7 @@ pub(crate) struct McpConnectionManager {
     clients: HashMap<String, std::sync::Arc<McpClient>>,
 
     /// Fully qualified tool name -> tool instance.
-    tools: HashMap<String, Tool>,
+    tools: HashMap<String, ToolInfo>,
 }
 
 impl McpConnectionManager {
@@ -79,12 +113,27 @@ impl McpConnectionManager {
 
         // Launch all configured servers concurrently.
         let mut join_set = JoinSet::new();
+        let mut errors = ClientStartErrors::new();
 
         for (server_name, cfg) in mcp_servers {
-            // TODO: Verify server name: require `^[a-zA-Z0-9_-]+$`?
+            // Validate server name before spawning
+            if !is_valid_mcp_server_name(&server_name) {
+                let error = anyhow::anyhow!(
+                    "invalid server name '{}': must match pattern ^[a-zA-Z0-9_-]+$",
+                    server_name
+                );
+                errors.insert(server_name, error);
+                continue;
+            }
+
             join_set.spawn(async move {
                 let McpServerConfig { command, args, env } = cfg;
-                let client_res = McpClient::new_stdio_client(command, args, env).await;
+                let client_res = McpClient::new_stdio_client(
+                    command.into(),
+                    args.into_iter().map(OsString::from).collect(),
+                    env,
+                )
+                .await;
                 match client_res {
                     Ok(client) => {
                         // Initialize the client.
@@ -93,10 +142,14 @@ impl McpConnectionManager {
                                 experimental: None,
                                 roots: None,
                                 sampling: None,
+                                // https://modelcontextprotocol.io/specification/2025-06-18/client/elicitation#capabilities
+                                // indicates this should be an empty object.
+                                elicitation: Some(json!({})),
                             },
                             client_info: Implementation {
                                 name: "codex-mcp-client".to_owned(),
                                 version: env!("CARGO_PKG_VERSION").to_owned(),
+                                title: Some("Codex".into()),
                             },
                             protocol_version: mcp_types::MCP_SCHEMA_VERSION.to_owned(),
                         };
@@ -117,7 +170,6 @@ impl McpConnectionManager {
 
         let mut clients: HashMap<String, std::sync::Arc<McpClient>> =
             HashMap::with_capacity(join_set.len());
-        let mut errors = ClientStartErrors::new();
 
         while let Some(res) = join_set.join_next().await {
             let (server_name, client_res) = res?; // JoinError propagation
@@ -132,7 +184,9 @@ impl McpConnectionManager {
             }
         }
 
-        let tools = list_all_tools(&clients).await?;
+        let all_tools = list_all_tools(&clients).await?;
+
+        let tools = qualify_tools(all_tools);
 
         Ok((Self { clients, tools }, errors))
     }
@@ -140,7 +194,10 @@ impl McpConnectionManager {
     /// Returns a single map that contains **all** tools. Each key is the
     /// fully-qualified name for the tool.
     pub fn list_all_tools(&self) -> HashMap<String, Tool> {
-        self.tools.clone()
+        self.tools
+            .iter()
+            .map(|(name, tool)| (name.clone(), tool.tool.clone()))
+            .collect()
     }
 
     /// Invoke the tool indicated by the (server, tool) pair.
@@ -162,13 +219,19 @@ impl McpConnectionManager {
             .await
             .with_context(|| format!("tool call failed for `{server}/{tool}`"))
     }
+
+    pub fn parse_tool_name(&self, tool_name: &str) -> Option<(String, String)> {
+        self.tools
+            .get(tool_name)
+            .map(|tool| (tool.server_name.clone(), tool.tool_name.clone()))
+    }
 }
 
 /// Query every server for its available tools and return a single map that
 /// contains **all** tools. Each key is the fully-qualified name for the tool.
-pub async fn list_all_tools(
+async fn list_all_tools(
     clients: &HashMap<String, std::sync::Arc<McpClient>>,
-) -> Result<HashMap<String, Tool>> {
+) -> Result<Vec<ToolInfo>> {
     let mut join_set = JoinSet::new();
 
     // Spawn one task per server so we can query them concurrently. This
@@ -185,18 +248,19 @@ pub async fn list_all_tools(
         });
     }
 
-    let mut aggregated: HashMap<String, Tool> = HashMap::with_capacity(join_set.len());
+    let mut aggregated: Vec<ToolInfo> = Vec::with_capacity(join_set.len());
 
     while let Some(join_res) = join_set.join_next().await {
         let (server_name, list_result) = join_res?;
         let list_result = list_result?;
 
         for tool in list_result.tools {
-            // TODO(mbolin): escape tool names that contain invalid characters.
-            let fq_name = fully_qualified_tool_name(&server_name, &tool.name);
-            if aggregated.insert(fq_name.clone(), tool).is_some() {
-                panic!("tool name collision for '{fq_name}': suspicious");
-            }
+            let tool_info = ToolInfo {
+                server_name: server_name.clone(),
+                tool_name: tool.name.clone(),
+                tool,
+            };
+            aggregated.push(tool_info);
         }
     }
 
@@ -208,3 +272,99 @@ pub async fn list_all_tools(
 
     Ok(aggregated)
 }
+
+fn is_valid_mcp_server_name(server_name: &str) -> bool {
+    !server_name.is_empty()
+        && server_name
+            .chars()
+            .all(|c| c.is_ascii_alphanumeric() || c == '_' || c == '-')
+}
+
+#[cfg(test)]
+#[allow(clippy::unwrap_used)]
+mod tests {
+    use super::*;
+    use mcp_types::ToolInputSchema;
+
+    fn create_test_tool(server_name: &str, tool_name: &str) -> ToolInfo {
+        ToolInfo {
+            server_name: server_name.to_string(),
+            tool_name: tool_name.to_string(),
+            tool: Tool {
+                annotations: None,
+                description: Some(format!("Test tool: {tool_name}")),
+                input_schema: ToolInputSchema {
+                    properties: None,
+                    required: None,
+                    r#type: "object".to_string(),
+                },
+                name: tool_name.to_string(),
+                output_schema: None,
+                title: None,
+            },
+        }
+    }
+
+    #[test]
+    fn test_qualify_tools_short_non_duplicated_names() {
+        let tools = vec![
+            create_test_tool("server1", "tool1"),
+            create_test_tool("server1", "tool2"),
+        ];
+
+        let qualified_tools = qualify_tools(tools);
+
+        assert_eq!(qualified_tools.len(), 2);
+        assert!(qualified_tools.contains_key("server1__tool1"));
+        assert!(qualified_tools.contains_key("server1__tool2"));
+    }
+
+    #[test]
+    fn test_qualify_tools_duplicated_names_skipped() {
+        let tools = vec![
+            create_test_tool("server1", "duplicate_tool"),
+            create_test_tool("server1", "duplicate_tool"),
+        ];
+
+        let qualified_tools = qualify_tools(tools);
+
+        // Only the first tool should remain, the second is skipped
+        assert_eq!(qualified_tools.len(), 1);
+        assert!(qualified_tools.contains_key("server1__duplicate_tool"));
+    }
+
+    #[test]
+    fn test_qualify_tools_long_names_same_server() {
+        let server_name = "my_server";
+
+        let tools = vec![
+            create_test_tool(
+                server_name,
+                "extremely_lengthy_function_name_that_absolutely_surpasses_all_reasonable_limits",
+            ),
+            create_test_tool(
+                server_name,
+                "yet_another_extremely_lengthy_function_name_that_absolutely_surpasses_all_reasonable_limits",
+            ),
+        ];
+
+        let qualified_tools = qualify_tools(tools);
+
+        assert_eq!(qualified_tools.len(), 2);
+
+        let mut keys: Vec<_> = qualified_tools.keys().cloned().collect();
+        keys.sort();
+
+        assert_eq!(keys[0].len(), 64);
+        assert_eq!(
+            keys[0],
+            "my_server__extremely_lena02e507efc5a9de88637e436690364fd4219e4ef"
+        );
+
+        assert_eq!(keys[1].len(), 64);
+        assert_eq!(
+            keys[1],
+            "my_server__yet_another_e1c3987bd9c50b826cbe1687966f79f0c602d19ca"
+        );
+    }
+}

codex-rs/core/src/mcp_tool_call.rs

@@ -1,4 +1,5 @@
 use std::time::Duration;
+use std::time::Instant;
 
 use tracing::error;
 
@@ -7,6 +8,7 @@ use crate::models::FunctionCallOutputPayload;
 use crate::models::ResponseInputItem;
 use crate::protocol::Event;
 use crate::protocol::EventMsg;
+use crate::protocol::McpInvocation;
 use crate::protocol::McpToolCallBeginEvent;
 use crate::protocol::McpToolCallEndEvent;
 
@@ -41,21 +43,28 @@ pub(crate) async fn handle_mcp_tool_call(
         }
     };
 
-    let tool_call_begin_event = EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
-        call_id: call_id.clone(),
+    let invocation = McpInvocation {
         server: server.clone(),
         tool: tool_name.clone(),
         arguments: arguments_value.clone(),
+    };
+
+    let tool_call_begin_event = EventMsg::McpToolCallBegin(McpToolCallBeginEvent {
+        call_id: call_id.clone(),
+        invocation: invocation.clone(),
     });
     notify_mcp_tool_call_event(sess, sub_id, tool_call_begin_event).await;
 
+    let start = Instant::now();
     // Perform the tool call.
     let result = sess
-        .call_tool(&server, &tool_name, arguments_value, timeout)
+        .call_tool(&server, &tool_name, arguments_value.clone(), timeout)
         .await
         .map_err(|e| format!("tool call error: {e}"));
     let tool_call_end_event = EventMsg::McpToolCallEnd(McpToolCallEndEvent {
         call_id: call_id.clone(),
+        invocation,
+        duration: start.elapsed(),
         result: result.clone(),
     });
 

codex-rs/core/src/model_provider_info.rs

@@ -9,13 +9,16 @@ use serde::Deserialize;
 use serde::Serialize;
 use std::collections::HashMap;
 use std::env::VarError;
+use std::time::Duration;
 
 use crate::error::EnvVarError;
-use crate::openai_api_key::get_openai_api_key;
 
 /// Value for the `OpenAI-Originator` header that is sent with requests to
 /// OpenAI.
 const OPENAI_ORIGINATOR_HEADER: &str = "codex_cli_rs";
+const DEFAULT_STREAM_IDLE_TIMEOUT_MS: u64 = 300_000;
+const DEFAULT_STREAM_MAX_RETRIES: u64 = 10;
+const DEFAULT_REQUEST_MAX_RETRIES: u64 = 4;
 
 /// Wire protocol that the provider speaks. Most third-party services only
 /// implement the classic OpenAI Chat Completions JSON schema, whereas OpenAI
@@ -26,7 +29,7 @@ const OPENAI_ORIGINATOR_HEADER: &str = "codex_cli_rs";
 #[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub enum WireApi {
-    /// The experimental “Responses” API exposed by OpenAI at `/v1/responses`.
+    /// The Responses API exposed by OpenAI at `/v1/responses`.
     Responses,
 
     /// Regular Chat Completions compatible with `/v1/chat/completions`.
@@ -40,7 +43,7 @@ pub struct ModelProviderInfo {
     /// Friendly display name.
     pub name: String,
     /// Base URL for the provider's OpenAI-compatible API.
-    pub base_url: String,
+    pub base_url: Option<String>,
     /// Environment variable that stores the user's API key for this provider.
     pub env_key: Option<String>,
 
@@ -64,6 +67,20 @@ pub struct ModelProviderInfo {
     /// value should be used. If the environment variable is not set, or the
     /// value is empty, the header will not be included in the request.
     pub env_http_headers: Option<HashMap<String, String>>,
+
+    /// Maximum number of times to retry a failed HTTP request to this provider.
+    pub request_max_retries: Option<u64>,
+
+    /// Number of times to retry reconnecting a dropped streaming response before failing.
+    pub stream_max_retries: Option<u64>,
+
+    /// Idle timeout (in milliseconds) to wait for activity on a streaming response before treating
+    /// the connection as lost.
+    pub stream_idle_timeout_ms: Option<u64>,
+
+    /// Whether this provider requires some form of standard authentication (API key, ChatGPT token).
+    #[serde(default)]
+    pub requires_auth: bool,
 }
 
 impl ModelProviderInfo {
@@ -79,11 +96,11 @@ impl ModelProviderInfo {
         &'a self,
         client: &'a reqwest::Client,
     ) -> crate::error::Result<reqwest::RequestBuilder> {
-        let api_key = self.api_key()?;
-
         let url = self.get_full_url();
 
         let mut builder = client.post(url);
+
+        let api_key = self.api_key()?;
         if let Some(key) = api_key {
             builder = builder.bearer_auth(key);
         }
@@ -103,9 +120,15 @@ impl ModelProviderInfo {
                     .join("&");
                 format!("?{full_params}")
             });
-        let base_url = &self.base_url;
+        let base_url = self
+            .base_url
+            .clone()
+            .unwrap_or("https://api.openai.com/v1".to_string());
+
         match self.wire_api {
-            WireApi::Responses => format!("{base_url}/responses{query_string}"),
+            WireApi::Responses => {
+                format!("{base_url}/responses{query_string}")
+            }
             WireApi::Chat => format!("{base_url}/chat/completions{query_string}"),
         }
     }
@@ -113,7 +136,10 @@ impl ModelProviderInfo {
     /// Apply provider-specific HTTP headers (both static and environment-based)
     /// onto an existing `reqwest::RequestBuilder` and return the updated
     /// builder.
-    fn apply_http_headers(&self, mut builder: reqwest::RequestBuilder) -> reqwest::RequestBuilder {
+    pub fn apply_http_headers(
+        &self,
+        mut builder: reqwest::RequestBuilder,
+    ) -> reqwest::RequestBuilder {
         if let Some(extra) = &self.http_headers {
             for (k, v) in extra {
                 builder = builder.header(k, v);
@@ -138,11 +164,7 @@ impl ModelProviderInfo {
     fn api_key(&self) -> crate::error::Result<Option<String>> {
         match &self.env_key {
             Some(env_key) => {
-                let env_value = if env_key == crate::openai_api_key::OPENAI_API_KEY_ENV_VAR {
-                    get_openai_api_key().map_or_else(|| Err(VarError::NotPresent), Ok)
-                } else {
-                    std::env::var(env_key)
-                };
+                let env_value = std::env::var(env_key);
                 env_value
                     .and_then(|v| {
                         if v.trim().is_empty() {
@@ -161,6 +183,25 @@ impl ModelProviderInfo {
             None => Ok(None),
         }
     }
+
+    /// Effective maximum number of request retries for this provider.
+    pub fn request_max_retries(&self) -> u64 {
+        self.request_max_retries
+            .unwrap_or(DEFAULT_REQUEST_MAX_RETRIES)
+    }
+
+    /// Effective maximum number of stream reconnection attempts for this provider.
+    pub fn stream_max_retries(&self) -> u64 {
+        self.stream_max_retries
+            .unwrap_or(DEFAULT_STREAM_MAX_RETRIES)
+    }
+
+    /// Effective idle timeout for streaming responses.
+    pub fn stream_idle_timeout(&self) -> Duration {
+        self.stream_idle_timeout_ms
+            .map(Duration::from_millis)
+            .unwrap_or(Duration::from_millis(DEFAULT_STREAM_IDLE_TIMEOUT_MS))
+    }
 }
 
 /// Built-in default provider list.
@@ -171,43 +212,51 @@ pub fn built_in_model_providers() -> HashMap<String, ModelProviderInfo> {
     // providers are bundled with Codex CLI, so we only include the OpenAI
     // provider by default. Users are encouraged to add to `model_providers`
     // in config.toml to add their own providers.
-    [
-        (
-            "openai",
-            P {
-                name: "OpenAI".into(),
-                // Allow users to override the default OpenAI endpoint by
-                // exporting `OPENAI_BASE_URL`. This is useful when pointing
-                // Codex at a proxy, mock server, or Azure-style deployment
-                // without requiring a full TOML override for the built-in
-                // OpenAI provider.
-                base_url: std::env::var("OPENAI_BASE_URL")
-                    .ok()
-                    .filter(|v| !v.trim().is_empty())
-                    .unwrap_or_else(|| "https://api.openai.com/v1".to_string()),
-                env_key: Some("OPENAI_API_KEY".into()),
-                env_key_instructions: Some("Create an API key (https://platform.openai.com) and export it as an environment variable.".into()),
-                wire_api: WireApi::Responses,
-                query_params: None,
-                http_headers: Some(
-                    [
-                        ("originator".to_string(), OPENAI_ORIGINATOR_HEADER.to_string()),
-                        ("version".to_string(), env!("CARGO_PKG_VERSION").to_string()),
-                    ]
-                        .into_iter()
-                        .collect(),
-                ),
-                env_http_headers: Some(
-                    [
-                        ("OpenAI-Organization".to_string(), "OPENAI_ORGANIZATION".to_string()),
-                        ("OpenAI-Project".to_string(), "OPENAI_PROJECT".to_string()),
-                    ]
-                        .into_iter()
-                        .collect(),
-                ),
-            },
-        ),
-    ]
+    [(
+        "openai",
+        P {
+            name: "OpenAI".into(),
+            // Allow users to override the default OpenAI endpoint by
+            // exporting `OPENAI_BASE_URL`. This is useful when pointing
+            // Codex at a proxy, mock server, or Azure-style deployment
+            // without requiring a full TOML override for the built-in
+            // OpenAI provider.
+            base_url: std::env::var("OPENAI_BASE_URL")
+                .ok()
+                .filter(|v| !v.trim().is_empty()),
+            env_key: None,
+            env_key_instructions: None,
+            wire_api: WireApi::Responses,
+            query_params: None,
+            http_headers: Some(
+                [
+                    (
+                        "originator".to_string(),
+                        OPENAI_ORIGINATOR_HEADER.to_string(),
+                    ),
+                    ("version".to_string(), env!("CARGO_PKG_VERSION").to_string()),
+                ]
+                .into_iter()
+                .collect(),
+            ),
+            env_http_headers: Some(
+                [
+                    (
+                        "OpenAI-Organization".to_string(),
+                        "OPENAI_ORGANIZATION".to_string(),
+                    ),
+                    ("OpenAI-Project".to_string(), "OPENAI_PROJECT".to_string()),
+                ]
+                .into_iter()
+                .collect(),
+            ),
+            // Use global defaults for retry/timeout unless overridden in config.toml.
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
+            requires_auth: true,
+        },
+    )]
     .into_iter()
     .map(|(k, v)| (k.to_string(), v))
     .collect()
@@ -227,13 +276,17 @@ base_url = "http://localhost:11434/v1"
         "#;
         let expected_provider = ModelProviderInfo {
             name: "Ollama".into(),
-            base_url: "http://localhost:11434/v1".into(),
+            base_url: Some("http://localhost:11434/v1".into()),
             env_key: None,
             env_key_instructions: None,
             wire_api: WireApi::Chat,
             query_params: None,
             http_headers: None,
             env_http_headers: None,
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
+            requires_auth: false,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
@@ -250,7 +303,7 @@ query_params = { api-version = "2025-04-01-preview" }
         "#;
         let expected_provider = ModelProviderInfo {
             name: "Azure".into(),
-            base_url: "https://xxxxx.openai.azure.com/openai".into(),
+            base_url: Some("https://xxxxx.openai.azure.com/openai".into()),
             env_key: Some("AZURE_OPENAI_API_KEY".into()),
             env_key_instructions: None,
             wire_api: WireApi::Chat,
@@ -259,6 +312,10 @@ query_params = { api-version = "2025-04-01-preview" }
             }),
             http_headers: None,
             env_http_headers: None,
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
+            requires_auth: false,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();
@@ -276,7 +333,7 @@ env_http_headers = { "X-Example-Env-Header" = "EXAMPLE_ENV_VAR" }
         "#;
         let expected_provider = ModelProviderInfo {
             name: "Example".into(),
-            base_url: "https://example.com".into(),
+            base_url: Some("https://example.com".into()),
             env_key: Some("API_KEY".into()),
             env_key_instructions: None,
             wire_api: WireApi::Chat,
@@ -287,6 +344,10 @@ env_http_headers = { "X-Example-Env-Header" = "EXAMPLE_ENV_VAR" }
             env_http_headers: Some(maplit::hashmap! {
                 "X-Example-Env-Header".to_string() => "EXAMPLE_ENV_VAR".to_string(),
             }),
+            request_max_retries: None,
+            stream_max_retries: None,
+            stream_idle_timeout_ms: None,
+            requires_auth: false,
         };
 
         let provider: ModelProviderInfo = toml::from_str(azure_provider_toml).unwrap();

codex-rs/core/src/models.rs

@@ -3,6 +3,7 @@ use std::collections::HashMap;
 use base64::Engine;
 use mcp_types::CallToolResult;
 use serde::Deserialize;
+use serde::Deserializer;
 use serde::Serialize;
 use serde::ser::Serializer;
 
@@ -37,6 +38,7 @@ pub enum ContentItem {
 #[serde(tag = "type", rename_all = "snake_case")]
 pub enum ResponseItem {
     Message {
+        id: Option<String>,
         role: String,
         content: Vec<ContentItem>,
     },
@@ -45,6 +47,7 @@ pub enum ResponseItem {
         summary: Vec<ReasoningItemReasoningSummary>,
         #[serde(default)]
         content: Vec<ReasoningItemContent>,
+        encrypted_content: Option<String>,
     },
     LocalShellCall {
         /// Set when using the chat completions API.
@@ -55,6 +58,7 @@ pub enum ResponseItem {
         action: LocalShellAction,
     },
     FunctionCall {
+        id: Option<String>,
         name: String,
         // The Responses API returns the function call arguments as a *string* that contains
         // JSON, not as an already‑parsed object. We keep it as a raw string here and let
@@ -80,7 +84,11 @@ pub enum ResponseItem {
 impl From<ResponseInputItem> for ResponseItem {
     fn from(item: ResponseInputItem) -> Self {
         match item {
-            ResponseInputItem::Message { role, content } => Self::Message { role, content },
+            ResponseInputItem::Message { role, content } => Self::Message {
+                role,
+                content,
+                id: None,
+            },
             ResponseInputItem::FunctionCallOutput { call_id, output } => {
                 Self::FunctionCallOutput { call_id, output }
             }
@@ -185,7 +193,7 @@ pub struct ShellToolCallParams {
     pub timeout_ms: Option<u64>,
 }
 
-#[derive(Deserialize, Debug, Clone)]
+#[derive(Debug, Clone)]
 pub struct FunctionCallOutputPayload {
     pub content: String,
     #[expect(dead_code)]
@@ -213,6 +221,19 @@ impl Serialize for FunctionCallOutputPayload {
     }
 }
 
+impl<'de> Deserialize<'de> for FunctionCallOutputPayload {
+    fn deserialize<D>(deserializer: D) -> Result<Self, D::Error>
+    where
+        D: Deserializer<'de>,
+    {
+        let s = String::deserialize(deserializer)?;
+        Ok(FunctionCallOutputPayload {
+            content: s,
+            success: None,
+        })
+    }
+}
+
 // Implement Display so callers can treat the payload like a plain string when logging or doing
 // trivial substring checks in tests (existing tests call `.contains()` on the output). Display
 // returns the raw `content` field.

codex-rs/core/src/openai_api_key.rs

@@ -1,24 +0,0 @@
-use std::env;
-use std::sync::LazyLock;
-use std::sync::RwLock;
-
-pub const OPENAI_API_KEY_ENV_VAR: &str = "OPENAI_API_KEY";
-
-static OPENAI_API_KEY: LazyLock<RwLock<Option<String>>> = LazyLock::new(|| {
-    let val = env::var(OPENAI_API_KEY_ENV_VAR)
-        .ok()
-        .and_then(|s| if s.is_empty() { None } else { Some(s) });
-    RwLock::new(val)
-});
-
-pub fn get_openai_api_key() -> Option<String> {
-    #![allow(clippy::unwrap_used)]
-    OPENAI_API_KEY.read().unwrap().clone()
-}
-
-pub fn set_openai_api_key(value: String) {
-    #![allow(clippy::unwrap_used)]
-    if !value.is_empty() {
-        *OPENAI_API_KEY.write().unwrap() = Some(value);
-    }
-}

codex-rs/core/src/openai_tools.rs

@@ -4,13 +4,14 @@ use std::collections::BTreeMap;
 use std::sync::LazyLock;
 
 use crate::client_common::Prompt;
+use crate::plan_tool::PLAN_TOOL;
 
 #[derive(Debug, Clone, Serialize)]
 pub(crate) struct ResponsesApiTool {
-    name: &'static str,
-    description: &'static str,
-    strict: bool,
-    parameters: JsonSchema,
+    pub(crate) name: &'static str,
+    pub(crate) description: &'static str,
+    pub(crate) strict: bool,
+    pub(crate) parameters: JsonSchema,
 }
 
 /// When serialized as JSON, this produces a valid "Tool" in the OpenAI
@@ -74,6 +75,7 @@ static DEFAULT_CODEX_MODEL_TOOLS: LazyLock<Vec<OpenAiTool>> =
 pub(crate) fn create_tools_json_for_responses_api(
     prompt: &Prompt,
     model: &str,
+    include_plan_tool: bool,
 ) -> crate::error::Result<Vec<serde_json::Value>> {
     // Assemble tool list: built-in tools + any extra tools from the prompt.
     let default_tools = if model.starts_with("codex") {
@@ -93,6 +95,10 @@ pub(crate) fn create_tools_json_for_responses_api(
             .map(|(name, tool)| mcp_tool_to_openai_tool(name, tool)),
     );
 
+    if include_plan_tool {
+        tools_json.push(serde_json::to_value(PLAN_TOOL.clone())?);
+    }
+
     Ok(tools_json)
 }
 
@@ -102,10 +108,12 @@ pub(crate) fn create_tools_json_for_responses_api(
 pub(crate) fn create_tools_json_for_chat_completions_api(
     prompt: &Prompt,
     model: &str,
+    include_plan_tool: bool,
 ) -> crate::error::Result<Vec<serde_json::Value>> {
     // We start with the JSON for the Responses API and than rewrite it to match
     // the chat completions tool call format.
-    let responses_api_tools_json = create_tools_json_for_responses_api(prompt, model)?;
+    let responses_api_tools_json =
+        create_tools_json_for_responses_api(prompt, model, include_plan_tool)?;
     let tools_json = responses_api_tools_json
         .into_iter()
         .filter_map(|mut tool| {

codex-rs/core/src/plan_tool.rs

@@ -0,0 +1,126 @@
+use std::collections::BTreeMap;
+use std::sync::LazyLock;
+
+use serde::Deserialize;
+use serde::Serialize;
+
+use crate::codex::Session;
+use crate::models::FunctionCallOutputPayload;
+use crate::models::ResponseInputItem;
+use crate::openai_tools::JsonSchema;
+use crate::openai_tools::OpenAiTool;
+use crate::openai_tools::ResponsesApiTool;
+use crate::protocol::Event;
+use crate::protocol::EventMsg;
+
+// Types for the TODO tool arguments matching codex-vscode/todo-mcp/src/main.rs
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum StepStatus {
+    Pending,
+    InProgress,
+    Completed,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct PlanItemArg {
+    pub step: String,
+    pub status: StepStatus,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct UpdatePlanArgs {
+    #[serde(default)]
+    pub explanation: Option<String>,
+    pub plan: Vec<PlanItemArg>,
+}
+
+pub(crate) static PLAN_TOOL: LazyLock<OpenAiTool> = LazyLock::new(|| {
+    let mut plan_item_props = BTreeMap::new();
+    plan_item_props.insert("step".to_string(), JsonSchema::String);
+    plan_item_props.insert("status".to_string(), JsonSchema::String);
+
+    let plan_items_schema = JsonSchema::Array {
+        items: Box::new(JsonSchema::Object {
+            properties: plan_item_props,
+            required: &["step", "status"],
+            additional_properties: false,
+        }),
+    };
+
+    let mut properties = BTreeMap::new();
+    properties.insert("explanation".to_string(), JsonSchema::String);
+    properties.insert("plan".to_string(), plan_items_schema);
+
+    OpenAiTool::Function(ResponsesApiTool {
+        name: "update_plan",
+        description: r#"Use the update_plan tool to keep the user updated on the current plan for the task.
+After understanding the user's task, call the update_plan tool with an initial plan. An example of a plan:
+1. Explore the codebase to find relevant files (status: in_progress)
+2. Implement the feature in the XYZ component (status: pending)
+3. Commit changes and make a pull request (status: pending)
+Each step should be a short, 1-sentence description.
+Until all the steps are finished, there should always be exactly one in_progress step in the plan.
+Call the update_plan tool whenever you finish a step, marking the completed step as `completed` and marking the next step as `in_progress`.
+Before running a command, consider whether or not you have completed the previous step, and make sure to mark it as completed before moving on to the next step.
+Sometimes, you may need to change plans in the middle of a task: call `update_plan` with the updated plan and make sure to provide an `explanation` of the rationale when doing so.
+When all steps are completed, call update_plan one last time with all steps marked as `completed`."#,
+        strict: false,
+        parameters: JsonSchema::Object {
+            properties,
+            required: &["plan"],
+            additional_properties: false,
+        },
+    })
+});
+
+/// This function doesn't do anything useful. However, it gives the model a structured way to record its plan that clients can read and render.
+/// So it's the _inputs_ to this function that are useful to clients, not the outputs and neither are actually useful for the model other
+/// than forcing it to come up and document a plan (TBD how that affects performance).
+pub(crate) async fn handle_update_plan(
+    session: &Session,
+    arguments: String,
+    sub_id: String,
+    call_id: String,
+) -> ResponseInputItem {
+    match parse_update_plan_arguments(arguments, &call_id) {
+        Ok(args) => {
+            let output = ResponseInputItem::FunctionCallOutput {
+                call_id,
+                output: FunctionCallOutputPayload {
+                    content: "Plan updated".to_string(),
+                    success: Some(true),
+                },
+            };
+            session
+                .send_event(Event {
+                    id: sub_id.to_string(),
+                    msg: EventMsg::PlanUpdate(args),
+                })
+                .await;
+            output
+        }
+        Err(output) => *output,
+    }
+}
+
+fn parse_update_plan_arguments(
+    arguments: String,
+    call_id: &str,
+) -> Result<UpdatePlanArgs, Box<ResponseInputItem>> {
+    match serde_json::from_str::<UpdatePlanArgs>(&arguments) {
+        Ok(args) => Ok(args),
+        Err(e) => {
+            let output = ResponseInputItem::FunctionCallOutput {
+                call_id: call_id.to_string(),
+                output: FunctionCallOutputPayload {
+                    content: format!("failed to parse function arguments: {e}"),
+                    success: None,
+                },
+            };
+            Err(Box::new(output))
+        }
+    }
+}

codex-rs/core/src/project_doc.rs

@@ -27,16 +27,16 @@ const PROJECT_DOC_SEPARATOR: &str = "\n\n--- project-doc ---\n\n";
 /// string of instructions.
 pub(crate) async fn get_user_instructions(config: &Config) -> Option<String> {
     match find_project_doc(config).await {
-        Ok(Some(project_doc)) => match &config.instructions {
+        Ok(Some(project_doc)) => match &config.user_instructions {
             Some(original_instructions) => Some(format!(
                 "{original_instructions}{PROJECT_DOC_SEPARATOR}{project_doc}"
             )),
             None => Some(project_doc),
         },
-        Ok(None) => config.instructions.clone(),
+        Ok(None) => config.user_instructions.clone(),
         Err(e) => {
             error!("error trying to find project doc: {e:#}");
-            config.instructions.clone()
+            config.user_instructions.clone()
         }
     }
 }
@@ -159,7 +159,7 @@ mod tests {
         config.cwd = root.path().to_path_buf();
         config.project_doc_max_bytes = limit;
 
-        config.instructions = instructions.map(ToOwned::to_owned);
+        config.user_instructions = instructions.map(ToOwned::to_owned);
         config
     }
 

codex-rs/core/src/protocol.rs

@@ -4,19 +4,23 @@
 //! between user and agent.
 
 use std::collections::HashMap;
+use std::fmt;
 use std::path::Path;
 use std::path::PathBuf;
 use std::str::FromStr;
+use std::time::Duration;
 
 use mcp_types::CallToolResult;
 use serde::Deserialize;
 use serde::Serialize;
+use strum_macros::Display;
 use uuid::Uuid;
 
 use crate::config_types::ReasoningEffort as ReasoningEffortConfig;
 use crate::config_types::ReasoningSummary as ReasoningSummaryConfig;
 use crate::message_history::HistoryEntry;
 use crate::model_provider_info::ModelProviderInfo;
+use crate::plan_tool::UpdatePlanArgs;
 
 /// Submission Queue Entry - requests from user
 #[derive(Debug, Clone, Deserialize, Serialize)]
@@ -44,8 +48,12 @@ pub enum Op {
         model_reasoning_effort: ReasoningEffortConfig,
         model_reasoning_summary: ReasoningSummaryConfig,
 
-        /// Model instructions
-        instructions: Option<String>,
+        /// Model instructions that are appended to the base instructions.
+        user_instructions: Option<String>,
+
+        /// Base instructions override.
+        base_instructions: Option<String>,
+
         /// When to escalate for approval for execution
         approval_policy: AskForApproval,
         /// How to sandbox commands executed in the system
@@ -69,6 +77,10 @@ pub enum Op {
         /// `ConfigureSession` operation so that the business-logic layer can
         /// operate deterministically.
         cwd: std::path::PathBuf,
+
+        /// Path to a rollout file to resume from.
+        #[serde(skip_serializing_if = "Option::is_none")]
+        resume_path: Option<std::path::PathBuf>,
     },
 
     /// Abort current task.
@@ -108,18 +120,23 @@ pub enum Op {
 
     /// Request a single history entry identified by `log_id` + `offset`.
     GetHistoryEntryRequest { offset: usize, log_id: u64 },
+
+    /// Request to shut down codex instance.
+    Shutdown,
 }
 
 /// Determines the conditions under which the user is consulted to approve
 /// running the command proposed by Codex.
-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash, Serialize, Deserialize)]
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Hash, Serialize, Deserialize, Display)]
 #[serde(rename_all = "kebab-case")]
+#[strum(serialize_all = "kebab-case")]
 pub enum AskForApproval {
     /// Under this policy, only "known safe" commands—as determined by
     /// `is_safe_command()`—that **only read files** are auto‑approved.
     /// Everything else will ask the user to approve.
     #[default]
     #[serde(rename = "untrusted")]
+    #[strum(serialize = "untrusted")]
     UnlessTrusted,
 
     /// *All* commands are auto‑approved, but they are expected to run inside a
@@ -263,8 +280,9 @@ pub struct Event {
 }
 
 /// Response event from the agent
-#[derive(Debug, Clone, Deserialize, Serialize)]
+#[derive(Debug, Clone, Deserialize, Serialize, Display)]
 #[serde(tag = "type", rename_all = "snake_case")]
+#[strum(serialize_all = "snake_case")]
 pub enum EventMsg {
     /// Error while executing a submission
     Error(ErrorEvent),
@@ -282,9 +300,15 @@ pub enum EventMsg {
     /// Agent text output message
     AgentMessage(AgentMessageEvent),
 
-    /// Reasoning summary from agent.
+    /// Agent text output delta message
+    AgentMessageDelta(AgentMessageDeltaEvent),
+
+    /// Reasoning event from agent.
     AgentReasoning(AgentReasoningEvent),
 
+    /// Agent reasoning delta event from agent.
+    AgentReasoningDelta(AgentReasoningDeltaEvent),
+  
     /// Raw chain-of-thought from agent.
     AgentReasoningContent(AgentReasoningContentEvent),
 
@@ -315,6 +339,11 @@ pub enum EventMsg {
 
     /// Response to GetHistoryEntryRequest.
     GetHistoryEntryResponse(GetHistoryEntryResponseEvent),
+
+    PlanUpdate(UpdatePlanArgs),
+
+    /// Notification that the agent is shutting down.
+    ShutdownComplete,
 }
 
 // Individual event payload types matching each `EventMsg` variant.
@@ -338,11 +367,46 @@ pub struct TokenUsage {
     pub total_tokens: u64,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct FinalOutput {
+    pub token_usage: TokenUsage,
+}
+
+impl From<TokenUsage> for FinalOutput {
+    fn from(token_usage: TokenUsage) -> Self {
+        Self { token_usage }
+    }
+}
+
+impl fmt::Display for FinalOutput {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        let u = &self.token_usage;
+        write!(
+            f,
+            "Token usage: total={} input={}{} output={}{}",
+            u.total_tokens,
+            u.input_tokens,
+            u.cached_input_tokens
+                .map(|c| format!(" (cached {c})"))
+                .unwrap_or_default(),
+            u.output_tokens,
+            u.reasoning_output_tokens
+                .map(|r| format!(" (reasoning {r})"))
+                .unwrap_or_default()
+        )
+    }
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct AgentMessageEvent {
     pub message: String,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct AgentMessageDeltaEvent {
+    pub delta: String,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct AgentReasoningEvent {
     pub text: String,
@@ -353,10 +417,12 @@ pub struct AgentReasoningContentEvent {
     pub text: String,
 }
 
+pub struct AgentReasoningDeltaEvent {
+    pub delta: String,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
-pub struct McpToolCallBeginEvent {
-    /// Identifier so this can be paired with the McpToolCallEnd event.
-    pub call_id: String,
+pub struct McpInvocation {
     /// Name of the MCP server as defined in the config.
     pub server: String,
     /// Name of the tool as given by the MCP server.
@@ -365,10 +431,19 @@ pub struct McpToolCallBeginEvent {
     pub arguments: Option<serde_json::Value>,
 }
 
+#[derive(Debug, Clone, Deserialize, Serialize)]
+pub struct McpToolCallBeginEvent {
+    /// Identifier so this can be paired with the McpToolCallEnd event.
+    pub call_id: String,
+    pub invocation: McpInvocation,
+}
+
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct McpToolCallEndEvent {
     /// Identifier for the corresponding McpToolCallBegin that finished.
     pub call_id: String,
+    pub invocation: McpInvocation,
+    pub duration: Duration,
     /// Result of the tool call. Note this could be an error.
     pub result: Result<CallToolResult, String>,
 }
@@ -406,6 +481,8 @@ pub struct ExecCommandEndEvent {
 
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct ExecApprovalRequestEvent {
+    /// Identifier for the associated exec call, if available.
+    pub call_id: String,
     /// The command to be executed.
     pub command: Vec<String>,
     /// The command's working directory.
@@ -417,6 +494,8 @@ pub struct ExecApprovalRequestEvent {
 
 #[derive(Debug, Clone, Deserialize, Serialize)]
 pub struct ApplyPatchApprovalRequestEvent {
+    /// Responses API call id for the associated patch apply call, if available.
+    pub call_id: String,
     pub changes: HashMap<PathBuf, FileChange>,
     /// Optional explanatory reason (e.g. request for extra write access).
     #[serde(skip_serializing_if = "Option::is_none")]

codex-rs/core/src/rollout.rs

@@ -1,33 +1,57 @@
-//! Functionality to persist a Codex conversation *rollout* – a linear list of
-//! [`ResponseItem`] objects exchanged during a session – to disk so that
-//! sessions can be replayed or inspected later (mirrors the behaviour of the
-//! upstream TypeScript implementation).
+//! Persist Codex session rollouts (.jsonl) so sessions can be replayed or inspected later.
 
 use std::fs::File;
 use std::fs::{self};
 use std::io::Error as IoError;
+use std::path::Path;
 
+use serde::Deserialize;
 use serde::Serialize;
+use serde_json::Value;
 use time::OffsetDateTime;
 use time::format_description::FormatItem;
 use time::macros::format_description;
 use tokio::io::AsyncWriteExt;
 use tokio::sync::mpsc::Sender;
 use tokio::sync::mpsc::{self};
+use tokio::sync::oneshot;
+use tracing::info;
+use tracing::warn;
 use uuid::Uuid;
 
 use crate::config::Config;
+use crate::git_info::GitInfo;
+use crate::git_info::collect_git_info;
 use crate::models::ResponseItem;
 
-/// Folder inside `~/.codex` that holds saved rollouts.
 const SESSIONS_SUBDIR: &str = "sessions";
 
+#[derive(Serialize, Deserialize, Clone, Default)]
+pub struct SessionMeta {
+    pub id: Uuid,
+    pub timestamp: String,
+    pub instructions: Option<String>,
+}
+
 #[derive(Serialize)]
-struct SessionMeta {
-    id: String,
-    timestamp: String,
+struct SessionMetaWithGit {
+    #[serde(flatten)]
+    meta: SessionMeta,
     #[serde(skip_serializing_if = "Option::is_none")]
-    instructions: Option<String>,
+    git: Option<GitInfo>,
+}
+
+#[derive(Serialize, Deserialize, Default, Clone)]
+pub struct SessionStateSnapshot {}
+
+#[derive(Serialize, Deserialize, Default, Clone)]
+pub struct SavedSession {
+    pub session: SessionMeta,
+    #[serde(default)]
+    pub items: Vec<ResponseItem>,
+    #[serde(default)]
+    pub state: SessionStateSnapshot,
+    pub session_id: Uuid,
 }
 
 /// Records all [`ResponseItem`]s for a session and flushes them to disk after
@@ -41,7 +65,13 @@ struct SessionMeta {
 /// ```
 #[derive(Clone)]
 pub(crate) struct RolloutRecorder {
-    tx: Sender<String>,
+    tx: Sender<RolloutCmd>,
+}
+
+enum RolloutCmd {
+    AddItems(Vec<ResponseItem>),
+    UpdateState(SessionStateSnapshot),
+    Shutdown { ack: oneshot::Sender<()> },
 }
 
 impl RolloutRecorder {
@@ -59,7 +89,6 @@ impl RolloutRecorder {
             timestamp,
         } = create_log_file(config, uuid)?;
 
-        // Build the static session metadata JSON first.
         let timestamp_format: &[FormatItem] = format_description!(
             "[year]-[month]-[day]T[hour]:[minute]:[second].[subsecond digits:3]Z"
         );
@@ -67,48 +96,33 @@ impl RolloutRecorder {
             .format(timestamp_format)
             .map_err(|e| IoError::other(format!("failed to format timestamp: {e}")))?;
 
-        let meta = SessionMeta {
-            timestamp,
-            id: session_id.to_string(),
-            instructions,
-        };
+        // Clone the cwd for the spawned task to collect git info asynchronously
+        let cwd = config.cwd.clone();
 
         // A reasonably-sized bounded channel. If the buffer fills up the send
         // future will yield, which is fine – we only need to ensure we do not
-        // perform *blocking* I/O on the caller’s thread.
-        let (tx, mut rx) = mpsc::channel::<String>(256);
+        // perform *blocking* I/O on the caller's thread.
+        let (tx, rx) = mpsc::channel::<RolloutCmd>(256);
 
         // Spawn a Tokio task that owns the file handle and performs async
         // writes. Using `tokio::fs::File` keeps everything on the async I/O
         // driver instead of blocking the runtime.
-        tokio::task::spawn(async move {
-            let mut file = tokio::fs::File::from_std(file);
+        tokio::task::spawn(rollout_writer(
+            tokio::fs::File::from_std(file),
+            rx,
+            Some(SessionMeta {
+                timestamp,
+                id: session_id,
+                instructions,
+            }),
+            cwd,
+        ));
 
-            while let Some(line) = rx.recv().await {
-                // Write line + newline, then flush to disk.
-                if let Err(e) = file.write_all(line.as_bytes()).await {
-                    tracing::warn!("rollout writer: failed to write line: {e}");
-                    break;
-                }
-                if let Err(e) = file.write_all(b"\n").await {
-                    tracing::warn!("rollout writer: failed to write newline: {e}");
-                    break;
-                }
-                if let Err(e) = file.flush().await {
-                    tracing::warn!("rollout writer: failed to flush: {e}");
-                    break;
-                }
-            }
-        });
-
-        let recorder = Self { tx };
-        // Ensure SessionMeta is the first item in the file.
-        recorder.record_item(&meta).await?;
-        Ok(recorder)
+        Ok(Self { tx })
     }
 
-    /// Append `items` to the rollout file.
     pub(crate) async fn record_items(&self, items: &[ResponseItem]) -> std::io::Result<()> {
+        let mut filtered = Vec::new();
         for item in items {
             match item {
                 // Note that function calls may look a bit strange if they are
@@ -117,27 +131,114 @@ impl RolloutRecorder {
                 ResponseItem::Message { .. }
                 | ResponseItem::LocalShellCall { .. }
                 | ResponseItem::FunctionCall { .. }
-                | ResponseItem::FunctionCallOutput { .. } => {}
-                ResponseItem::Reasoning { .. } | ResponseItem::Other => {
+                | ResponseItem::FunctionCallOutput { .. }
+                | ResponseItem::Reasoning { .. } => filtered.push(item.clone()),
+                ResponseItem::Other => {
                     // These should never be serialized.
                     continue;
                 }
             }
-            self.record_item(item).await?;
         }
-        Ok(())
+        if filtered.is_empty() {
+            return Ok(());
+        }
+        self.tx
+            .send(RolloutCmd::AddItems(filtered))
+            .await
+            .map_err(|e| IoError::other(format!("failed to queue rollout items: {e}")))
     }
 
-    async fn record_item(&self, item: &impl Serialize) -> std::io::Result<()> {
-        // Serialize the item to JSON first so that the writer thread only has
-        // to perform the actual write.
-        let json = serde_json::to_string(item)
-            .map_err(|e| IoError::other(format!("failed to serialize response items: {e}")))?;
-
+    pub(crate) async fn record_state(&self, state: SessionStateSnapshot) -> std::io::Result<()> {
         self.tx
-            .send(json)
+            .send(RolloutCmd::UpdateState(state))
             .await
-            .map_err(|e| IoError::other(format!("failed to queue rollout item: {e}")))
+            .map_err(|e| IoError::other(format!("failed to queue rollout state: {e}")))
+    }
+
+    pub async fn resume(
+        path: &Path,
+        cwd: std::path::PathBuf,
+    ) -> std::io::Result<(Self, SavedSession)> {
+        info!("Resuming rollout from {path:?}");
+        let text = tokio::fs::read_to_string(path).await?;
+        let mut lines = text.lines();
+        let meta_line = lines
+            .next()
+            .ok_or_else(|| IoError::other("empty session file"))?;
+        let session: SessionMeta = serde_json::from_str(meta_line)
+            .map_err(|e| IoError::other(format!("failed to parse session meta: {e}")))?;
+        let mut items = Vec::new();
+        let mut state = SessionStateSnapshot::default();
+
+        for line in lines {
+            if line.trim().is_empty() {
+                continue;
+            }
+            let v: Value = match serde_json::from_str(line) {
+                Ok(v) => v,
+                Err(_) => continue,
+            };
+            if v.get("record_type")
+                .and_then(|rt| rt.as_str())
+                .map(|s| s == "state")
+                .unwrap_or(false)
+            {
+                if let Ok(s) = serde_json::from_value::<SessionStateSnapshot>(v.clone()) {
+                    state = s
+                }
+                continue;
+            }
+            match serde_json::from_value::<ResponseItem>(v.clone()) {
+                Ok(item) => match item {
+                    ResponseItem::Message { .. }
+                    | ResponseItem::LocalShellCall { .. }
+                    | ResponseItem::FunctionCall { .. }
+                    | ResponseItem::FunctionCallOutput { .. }
+                    | ResponseItem::Reasoning { .. } => items.push(item),
+                    ResponseItem::Other => {}
+                },
+                Err(e) => {
+                    warn!("failed to parse item: {v:?}, error: {e}");
+                }
+            }
+        }
+
+        let saved = SavedSession {
+            session: session.clone(),
+            items: items.clone(),
+            state: state.clone(),
+            session_id: session.id,
+        };
+
+        let file = std::fs::OpenOptions::new()
+            .append(true)
+            .read(true)
+            .open(path)?;
+
+        let (tx, rx) = mpsc::channel::<RolloutCmd>(256);
+        tokio::task::spawn(rollout_writer(
+            tokio::fs::File::from_std(file),
+            rx,
+            None,
+            cwd,
+        ));
+        info!("Resumed rollout successfully from {path:?}");
+        Ok((Self { tx }, saved))
+    }
+
+    pub async fn shutdown(&self) -> std::io::Result<()> {
+        let (tx_done, rx_done) = oneshot::channel();
+        match self.tx.send(RolloutCmd::Shutdown { ack: tx_done }).await {
+            Ok(_) => rx_done
+                .await
+                .map_err(|e| IoError::other(format!("failed waiting for rollout shutdown: {e}"))),
+            Err(e) => {
+                warn!("failed to send rollout shutdown command: {e}");
+                Err(IoError::other(format!(
+                    "failed to send rollout shutdown command: {e}"
+                )))
+            }
+        }
     }
 }
 
@@ -153,13 +254,15 @@ struct LogFileInfo {
 }
 
 fn create_log_file(config: &Config, session_id: Uuid) -> std::io::Result<LogFileInfo> {
-    // Resolve ~/.codex/sessions and create it if missing.
-    let mut dir = config.codex_home.clone();
-    dir.push(SESSIONS_SUBDIR);
-    fs::create_dir_all(&dir)?;
-
+    // Resolve ~/.codex/sessions/YYYY/MM/DD and create it if missing.
     let timestamp = OffsetDateTime::now_local()
         .map_err(|e| IoError::other(format!("failed to get local time: {e}")))?;
+    let mut dir = config.codex_home.clone();
+    dir.push(SESSIONS_SUBDIR);
+    dir.push(timestamp.year().to_string());
+    dir.push(format!("{:02}", u8::from(timestamp.month())));
+    dir.push(format!("{:02}", timestamp.day()));
+    fs::create_dir_all(&dir)?;
 
     // Custom format for YYYY-MM-DDThh-mm-ss. Use `-` instead of `:` for
     // compatibility with filesystems that do not allow colons in filenames.
@@ -183,3 +286,77 @@ fn create_log_file(config: &Config, session_id: Uuid) -> std::io::Result<LogFile
         timestamp,
     })
 }
+
+async fn rollout_writer(
+    file: tokio::fs::File,
+    mut rx: mpsc::Receiver<RolloutCmd>,
+    mut meta: Option<SessionMeta>,
+    cwd: std::path::PathBuf,
+) -> std::io::Result<()> {
+    let mut writer = JsonlWriter { file };
+
+    // If we have a meta, collect git info asynchronously and write meta first
+    if let Some(session_meta) = meta.take() {
+        let git_info = collect_git_info(&cwd).await;
+        let session_meta_with_git = SessionMetaWithGit {
+            meta: session_meta,
+            git: git_info,
+        };
+
+        // Write the SessionMeta as the first item in the file
+        writer.write_line(&session_meta_with_git).await?;
+    }
+
+    // Process rollout commands
+    while let Some(cmd) = rx.recv().await {
+        match cmd {
+            RolloutCmd::AddItems(items) => {
+                for item in items {
+                    match item {
+                        ResponseItem::Message { .. }
+                        | ResponseItem::LocalShellCall { .. }
+                        | ResponseItem::FunctionCall { .. }
+                        | ResponseItem::FunctionCallOutput { .. }
+                        | ResponseItem::Reasoning { .. } => {
+                            writer.write_line(&item).await?;
+                        }
+                        ResponseItem::Other => {}
+                    }
+                }
+            }
+            RolloutCmd::UpdateState(state) => {
+                #[derive(Serialize)]
+                struct StateLine<'a> {
+                    record_type: &'static str,
+                    #[serde(flatten)]
+                    state: &'a SessionStateSnapshot,
+                }
+                writer
+                    .write_line(&StateLine {
+                        record_type: "state",
+                        state: &state,
+                    })
+                    .await?;
+            }
+            RolloutCmd::Shutdown { ack } => {
+                let _ = ack.send(());
+            }
+        }
+    }
+
+    Ok(())
+}
+
+struct JsonlWriter {
+    file: tokio::fs::File,
+}
+
+impl JsonlWriter {
+    async fn write_line(&mut self, item: &impl serde::Serialize) -> std::io::Result<()> {
+        let mut json = serde_json::to_string(item)?;
+        json.push('\n');
+        let _ = self.file.write_all(json.as_bytes()).await;
+        self.file.flush().await?;
+        Ok(())
+    }
+}

codex-rs/core/src/safety.rs

@@ -75,9 +75,6 @@ pub fn assess_command_safety(
     sandbox_policy: &SandboxPolicy,
     approved: &HashSet<Vec<String>>,
 ) -> SafetyCheck {
-    use AskForApproval::*;
-    use SandboxPolicy::*;
-
     // A command is "trusted" because either:
     // - it belongs to a set of commands we consider "safe" by default, or
     // - the user has explicitly approved the command for this session
@@ -97,6 +94,16 @@ pub fn assess_command_safety(
         };
     }
 
+    assess_safety_for_untrusted_command(approval_policy, sandbox_policy)
+}
+
+pub(crate) fn assess_safety_for_untrusted_command(
+    approval_policy: AskForApproval,
+    sandbox_policy: &SandboxPolicy,
+) -> SafetyCheck {
+    use AskForApproval::*;
+    use SandboxPolicy::*;
+
     match (approval_policy, sandbox_policy) {
         (UnlessTrusted, _) => {
             // Even though the user may have opted into DangerFullAccess,

codex-rs/core/src/shell.rs

@@ -0,0 +1,236 @@
+use shlex;
+
+#[derive(Debug, PartialEq, Eq)]
+pub struct ZshShell {
+    shell_path: String,
+    zshrc_path: String,
+}
+
+#[derive(Debug, PartialEq, Eq)]
+pub enum Shell {
+    Zsh(ZshShell),
+    Unknown,
+}
+
+impl Shell {
+    pub fn format_default_shell_invocation(&self, command: Vec<String>) -> Option<Vec<String>> {
+        match self {
+            Shell::Zsh(zsh) => {
+                if !std::path::Path::new(&zsh.zshrc_path).exists() {
+                    return None;
+                }
+
+                let mut result = vec![zsh.shell_path.clone()];
+                result.push("-lc".to_string());
+
+                let joined = strip_bash_lc(&command)
+                    .or_else(|| shlex::try_join(command.iter().map(|s| s.as_str())).ok());
+
+                if let Some(joined) = joined {
+                    result.push(format!("source {} && ({joined})", zsh.zshrc_path));
+                } else {
+                    return None;
+                }
+                Some(result)
+            }
+            Shell::Unknown => None,
+        }
+    }
+}
+
+fn strip_bash_lc(command: &Vec<String>) -> Option<String> {
+    match command.as_slice() {
+        // exactly three items
+        [first, second, third]
+            // first two must be "bash", "-lc"
+            if first == "bash" && second == "-lc" =>
+        {
+            Some(third.clone())
+        }
+        _ => None,
+    }
+}
+
+#[cfg(target_os = "macos")]
+pub async fn default_user_shell() -> Shell {
+    use tokio::process::Command;
+    use whoami;
+
+    let user = whoami::username();
+    let home = format!("/Users/{user}");
+    let output = Command::new("dscl")
+        .args([".", "-read", &home, "UserShell"])
+        .output()
+        .await
+        .ok();
+    match output {
+        Some(o) => {
+            if !o.status.success() {
+                return Shell::Unknown;
+            }
+            let stdout = String::from_utf8_lossy(&o.stdout);
+            for line in stdout.lines() {
+                if let Some(shell_path) = line.strip_prefix("UserShell: ") {
+                    if shell_path.ends_with("/zsh") {
+                        return Shell::Zsh(ZshShell {
+                            shell_path: shell_path.to_string(),
+                            zshrc_path: format!("{home}/.zshrc"),
+                        });
+                    }
+                }
+            }
+
+            Shell::Unknown
+        }
+        _ => Shell::Unknown,
+    }
+}
+
+#[cfg(not(target_os = "macos"))]
+pub async fn default_user_shell() -> Shell {
+    Shell::Unknown
+}
+
+#[cfg(test)]
+#[cfg(target_os = "macos")]
+mod tests {
+    use super::*;
+    use std::process::Command;
+
+    #[tokio::test]
+    #[expect(clippy::unwrap_used)]
+    async fn test_current_shell_detects_zsh() {
+        let shell = Command::new("sh")
+            .arg("-c")
+            .arg("echo $SHELL")
+            .output()
+            .unwrap();
+
+        let home = std::env::var("HOME").unwrap();
+        let shell_path = String::from_utf8_lossy(&shell.stdout).trim().to_string();
+        if shell_path.ends_with("/zsh") {
+            assert_eq!(
+                default_user_shell().await,
+                Shell::Zsh(ZshShell {
+                    shell_path: shell_path.to_string(),
+                    zshrc_path: format!("{home}/.zshrc",),
+                })
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn test_run_with_profile_zshrc_not_exists() {
+        let shell = Shell::Zsh(ZshShell {
+            shell_path: "/bin/zsh".to_string(),
+            zshrc_path: "/does/not/exist/.zshrc".to_string(),
+        });
+        let actual_cmd = shell.format_default_shell_invocation(vec!["myecho".to_string()]);
+        assert_eq!(actual_cmd, None);
+    }
+
+    #[expect(clippy::unwrap_used)]
+    #[tokio::test]
+    async fn test_run_with_profile_escaping_and_execution() {
+        let shell_path = "/bin/zsh";
+
+        let cases = vec![
+            (
+                vec!["myecho"],
+                vec![shell_path, "-lc", "source ZSHRC_PATH && (myecho)"],
+                Some("It works!\n"),
+            ),
+            (
+                vec!["myecho"],
+                vec![shell_path, "-lc", "source ZSHRC_PATH && (myecho)"],
+                Some("It works!\n"),
+            ),
+            (
+                vec!["bash", "-c", "echo 'single' \"double\""],
+                vec![
+                    shell_path,
+                    "-lc",
+                    "source ZSHRC_PATH && (bash -c \"echo 'single' \\\"double\\\"\")",
+                ],
+                Some("single double\n"),
+            ),
+            (
+                vec!["bash", "-lc", "echo 'single' \"double\""],
+                vec![
+                    shell_path,
+                    "-lc",
+                    "source ZSHRC_PATH && (echo 'single' \"double\")",
+                ],
+                Some("single double\n"),
+            ),
+        ];
+        for (input, expected_cmd, expected_output) in cases {
+            use std::collections::HashMap;
+            use std::path::PathBuf;
+            use std::sync::Arc;
+
+            use tokio::sync::Notify;
+
+            use crate::exec::ExecParams;
+            use crate::exec::SandboxType;
+            use crate::exec::process_exec_tool_call;
+            use crate::protocol::SandboxPolicy;
+
+            // create a temp directory with a zshrc file in it
+            let temp_home = tempfile::tempdir().unwrap();
+            let zshrc_path = temp_home.path().join(".zshrc");
+            std::fs::write(
+                &zshrc_path,
+                r#"
+                    set -x
+                    function myecho {
+                        echo 'It works!'
+                    }
+                    "#,
+            )
+            .unwrap();
+            let shell = Shell::Zsh(ZshShell {
+                shell_path: shell_path.to_string(),
+                zshrc_path: zshrc_path.to_str().unwrap().to_string(),
+            });
+
+            let actual_cmd = shell
+                .format_default_shell_invocation(input.iter().map(|s| s.to_string()).collect());
+            let expected_cmd = expected_cmd
+                .iter()
+                .map(|s| {
+                    s.replace("ZSHRC_PATH", zshrc_path.to_str().unwrap())
+                        .to_string()
+                })
+                .collect();
+
+            assert_eq!(actual_cmd, Some(expected_cmd));
+            // Actually run the command and check output/exit code
+            let output = process_exec_tool_call(
+                ExecParams {
+                    command: actual_cmd.unwrap(),
+                    cwd: PathBuf::from(temp_home.path()),
+                    timeout_ms: None,
+                    env: HashMap::from([(
+                        "HOME".to_string(),
+                        temp_home.path().to_str().unwrap().to_string(),
+                    )]),
+                },
+                SandboxType::None,
+                Arc::new(Notify::new()),
+                &SandboxPolicy::DangerFullAccess,
+                &None,
+            )
+            .await
+            .unwrap();
+
+            assert_eq!(output.exit_code, 0, "input: {input:?} output: {output:?}");
+            if let Some(expected) = expected_output {
+                assert_eq!(
+                    output.stdout, expected,
+                    "input: {input:?} output: {output:?}"
+                );
+            }
+        }
+    }
+}