[codex] fix(prompt): unify serial tool-calling instructions

[yzlu0917]

Summary

• replace the agent-mode base instruction that previously encouraged simultaneous read-only tool calls
• align the system-message tool codeblock framework with the actual serial tool-calling policy used in practice
• add regression tests covering both the default agent system message and generated tool instructions

Why

Continue currently gives models conflicting instructions about tool sequencing. The base agent system message says multiple read-only tools can be called simultaneously, while the system-message tool framework says only one tool can be called at a time. In practice, Continue already defaults to serial tool execution and disables parallel tool calls for OpenAI-compatible requests in the common path, so the prompt should match reality.

This removes the contradiction and makes the default guidance consistent: call one tool, wait for the result, then decide the next step.

Validation

• ran npm test -- src/redux/util/getBaseSystemMessage.test.ts in gui
• ran npm run vitest -- tools/systemMessageTools/toolCodeblocks/buildSystemMessage.vitest.ts in core
• ran git diff --check

Closes #9308

---

Summary by cubic

Unifies tool-calling guidance to a single serial policy so prompts match actual behavior. Aligns with #9308 and standardizes OpenRouter headers.

• Bug Fixes
  • Prompts: Update default agent system message and tool codeblock rules to require serial tool calls; remove “simultaneously” language.
  • Tests: Add regression tests in core and gui to assert serial instruction and absence of “simultaneously”.
  • OpenRouter: Use OPENROUTER_HEADERS from @continuedev/openai-adapters in core/llm/llms/OpenRouter; export headers and update keys to X-OpenRouter-Title and X-OpenRouter-Categories.

^{Written for commit 2cbd793. Summary will update on new commits.}

[cubic-dev-ai]

1 issue found across 7 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="extensions/vscode/src/util/runCommandInTerminal.ts">

<violation number="1" location="extensions/vscode/src/util/runCommandInTerminal.ts:71">
P1: Remote terminal creation can resolve to an unrelated newly opened terminal, causing command execution in the wrong terminal session.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[cubic-dev-ai]

P1: Remote terminal creation can resolve to an unrelated newly opened terminal, causing command execution in the wrong terminal session.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At extensions/vscode/src/util/runCommandInTerminal.ts, line 71:

<comment>Remote terminal creation can resolve to an unrelated newly opened terminal, causing command execution in the wrong terminal session.</comment>

<file context>
@@ -0,0 +1,118 @@
+
+      settled = true;
+      cleanup();
+      resolve(terminal);
+    });
+
</file context>

[chatgpt-codex-connector]

💡 Codex Review

continue/core/util/history.ts

Lines 42 to 44 in 6daeaa4

	if (options.workspaceDirectory) {
	const target = options.workspaceDirectory.toLowerCase();
	sessions = sessions.filter(

Apply workspace filter after merging remote sessions

This workspace filter only runs inside HistoryManager.list, which covers local sessions, but history/list in core/core.ts appends all remote sessions afterward. As a result, callers that pass workspaceDirectory still get unrelated remote sessions whenever remote history is enabled, so workspace-scoped history views are not actually scoped. Please apply the workspace filter after local and remote results are combined (or filter remote sessions before merging).

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".